Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

First results from the LUCID-Timepix spacecraft payload onboard the TechDemoSat-1 satellite in Low Earth Orbit

First results from the LUCID-Timepix spacecraft payload onboard the TechDemoSat-1 satellite in... The Langton Ultimate Cosmic ray Intensity Detector (LUCID) is a payload onboard the satellite TechDemoSat-1, used to study the radiation environment in Low Earth Orbit (635km). LUCID operated from 2014 to 2017, collecting over 2.1 million frames of radiation data from its ve Timepix detectors on board. LUCID is one of the rst uses of the Timepix detector technology in open space, with the data providing useful insight into the performance of this technology in new environments. It provides high-sensitivity imaging measurements of the mixed radiation eld, with a wide dynamic range in terms of spectral response, particle type and direction. The data has been analysed using computing resources provided by GridPP, with a new machine learning algorithm that uses the Tensor ow framework. This algorithm provides a new approach to processing Medipix data, using a training set of human labelled tracks, providing greater particle classi cation accuracy than other algorithms. For managing the LUCID data, we have developed an online platform called Timepix Analysis Platform at School (TAPAS). This provides a swift and simple way for users to analyse data that they collect using Timepix detectors from both LUCID and other experiments. We also present some possible future uses of the LUCID data and Medipix detectors in space. Keywords: Low Earth Orbit, Space Radiation, Trapped Radiation, Cosmic Rays, Timepix, South Atlantic Anomaly 1. Introduction of sources, from solar (typically E < 10 eV, including Solar Energetic Particles) to within the Milky Way Cosmic radiation consists of high energy particles 9 15 (typically, 10 eV < E < 10 eV Galactic Cosmic Rays), produced by a variety of extra-terrestrial sources. In 15 to extragalactic (typically 10 eV < E), with the source general, cosmic radiation falls into three categories 19 of the highest energy particles (10 eV < E) still heavily depending on their source; Galactic Cosmic Rays (GCRs) debated (see Greisen, 1966, Pierre Auger Collaboration which originate outside the solar system, Solar Energetic et al., 2017, Bell et al., 2018 and many more). See Ferrari Particles (SEPs) which come from the sun, and charged and Szuszkiewicz (2009), Blasi (2013), Deligny and O. particles that are trapped by the Earth's magnetic eld. (2016) and Amato and Blasi (2017) for recent reviews of the study of cosmic rays. When the particles are detected directly they are known as primary particles; particles produced by an Cosmic rays are of great scienti c interest for a variety interaction between a primary particle and some ob- of reasons. In general, (secondary) cosmic rays make up structing medium (e.g. hitting the Earth's atmosphere) around 10% of the background radiation we experience are known as secondary particles. Primary cosmic rays on earth (United Nations Scienti c Committee on the are made up of a variety of particles (protons, electrons, E ects of Atomic Radiation, 2008). High energy rays gammas, light nuclei); an even wider range of particles are have long been a probe of fundamental physics, from the typically produced in secondary particle showers, includ- famous example of the e ect of time dilation on muon ing neutrons, Minimum Ionising Particles (MIPS), which decay rate (Rossi and Hall, 1941, Frisch and Smith, 1963), are usually muons, and pions. Primary cosmic rays span to more recently the new physics suggested by a lack 4 20 a vast range of energies ( 10 10 eV - in contrast the of anisotropy in the cosmic ray electron-positron ratio - maximum energy reached by the Large Hadron Collider potentially from dark matter annihilation (Aguilar et al., is  10 eV, CMS Collaboration 2017) from a large range 2013). High-energy cosmic rays can give valuable infor- mation about high-energy astrophysics, e.g. shockwaves in supernovae (Giuliani et al., 2011, Ackermann et al., Corresponding author 2013), merging neutron stars (Komiya and Shigeyama, Email address: mail@willfurnell.com (Will Furnell) Preprint submitted to Advances in Space Research October 31, 2018 arXiv:1810.12876v1 [astro-ph.IM] 30 Oct 2018 2017) and potentially active galactic nuclei (The Pierre Four of these have been in near continuous operation since Auger Collaboration, 2007, Pierre Auger Collaboration 2012, operated via an onboard laptop. The second use et al., 2017). Cosmic rays can be viewed as complemen- of Medipix in space was on the European Space Agency tary messengers in multi-messenger astronomy, alongside (ESA) Proba V mission, launched on the 7th May 2013 photons, neutrinos and gravitational waves (Branchesi, to Low Earth Orbit (LEO) with an altitude of 820km, 2016). See Ginzburg (1996), Kotera and Olinto (2011) with the spacecraft payload Space Application of Timepix and Castellina and Donato (2013), for recent reviews Radiation Monitor (SATRAM, Granja et al., 2016) of the role of cosmic rays in astrophysics. More locally onboard. SATRAM carries a single Timepix detector and cosmic rays are a probe of solar physics (e.g. see Potgieter, is operating and continuously taking data today. In addi- 2013), and are a key component of `space weather' (e.g. tion, Exploration Flight Test 1 (EFT-1), the rst ight see Turner et al., 2014). Space weather can have an of the Orion Multi-Purpose Crew Vehicle (MPCV) on a impact both on ground based communications systems two orbit, 4.5 hour trip on the 5th December 2014 took (e.g. a modern day Carrington event, see Love et al., Medipix data the farthest from Earth to date at 5910 2016) satellites and spacecraft electronics (Choi et al., km (Gaza et al., 2017). Most recently, on the 23rd June 2011) and organisms (including humans). Understanding 2017 the cubesat VZLUSAT-1 (altitude 510km, D aniel the radiation environment is vital for understanding the et al., 2016, Urban et al., 2017) carried a miniaturised impact of the dose astronauts receive on the International x-ray telescope, that uses Timepix detectors (Baca et al., Space Station (ISS, e.g. the link between received 2016), into orbit for astrophysical, space weather studies, dose and susceptibility to cataracts, Cucinotta et al., and terrestrial X-ray monitoring applications (see Pina 2001, see Cucinotta, 2007 for an overview), or on a hy- et al., 2015). pothetical voyage to Mars (Kerr, 2013, Zeitlin et al., 2013). In this paper we report the rst results from one of The development of novel energetic particle detector the early uses of Medipix in orbit (and the rst on a technologies at the European Organization for Nuclear commercial platform, and the rst with Medipix detectors Research (CERN) provides an opportunity to improve in a 3D con guration), the Langton Ultimate Cosmic ray measurements of cosmic rays both in space and on the Intensity Detector (LUCID) on board TechDemoSat-1. ground. The turn of the 21st Century saw the advent of The dominant source of particles detected by LUCID are photon counting pixel detectors for radiation detection trapped electrons and protons, and the instrument is de- with the development of Medipix detectors (Bisogni et al., signed such that it could be deployed as a hosted payload 1998, Campbell et al., 1998). Medipix detectors (now for satellite environmental monitoring. We discuss the de- entering their fourth generation, Medipix1 Amendolia sign of the instrument and its operations, a new platform et al., 1999, Medipix2 Llopart et al., 2002, Medipix3 TAPAS for managing the large amounts of data produced Ballabriga et al., 2011, Medipix4 collaboration founded by the experiment and a new machine learning algorithm 2016) can detect and di erentiate between many types for the automated classi cation of particles which can be of ionizing radiation, and present many advantages used for other Medipix applications. The rst results and compared to other methods - but at the cost of a small some early science applications (e.g. mapping out the collecting area and comparatively high expense. Medipix South Atlantic Anomaly) are also presented. In addition, detectors have been used in a wide range of applications, LUCID is linked to a extensive programme of education including high-energy physics experiments (Grei enberg and research in the classroom, CERN@school, where et al., 2009, Vykydal et al., 2009, Collins et al., 2011), students can use Timepix detectors for both novel tests medical physics (hence `Medi', e.g. Blanchot et al., 2006, of traditional classroom experiments (e.g. inverse square Butzer et al., 2008, Marti s kov a et al., 2011, Jakubek law, see Whyntie and Parker, 2013) as well as original et al., 2011, Hartmann, 2013) and small animal imaging science, for example the Radiation In Soil Experiment (Accorsi et al., 2008). (RISE) which has measured the radiation in di erent geological samples across the UK, and construction of More recently, there has been increased interest in a robotic three-dimensional radiation scanner (Whyntie their application in space (Kroupa et al., 2015, Granja et al., 2016, see also Colthurst et al., 2015, Parker, 2017, et al., 2016, Gaza et al., 2017, Urban et al., 2017). In Parker et al., 2018 for further general information). All of particular their ability to distinguish between di erent these projects are managed through the same data storage particle types and give angular information have, for and reduction pipeline used in this work, see section 3.2. example, proved valuable in understanding the radiation Cubesats have already been shown to be highly e ective environment of the ISS. Using Medipix in space was rst educational tools e.g. Li et al. (2013), and other Medipix discussed in Pinsky et al. (2011). Seven NASA/IEAP- devices have been used in the CERN@school programme developed Radiation Environment Monitors (REMs), e.g. RasPIX Turecek (2016). Timepix detectors in compact USB mounting, have been deployed to the ISS (altitude 400km), see Turecek et al. In addition to LUCID, the ISS REMS, SATRAM (2011b), Kroupa et al. (2015) and Stoe et al. (2015). and the VZLUSAT-1 x-ray telescope, there are also 2 future planned deployments, see Pinsky et al. (2016). et al., 2002, Llopart et al., 2007; Ballabriga et al., 2011). Future missions include; a particle telescope architecture The detectors used are equipped with a 300m silicon containing two Timepix detectors in sync, on the Rapid sensor. The Timepix chips contain 256 256 pixels, each International Scienti c Experiment Satellite (RISESAT, measuring 55m on each side, giving a total collecting Kuwahara et al., 2011, Granja et al., 2014a), a Japanese area of 1.98cm . The circuitry to digitise the output of FIRST mission to orbit at 700km, further Medipix being each pixel is contained within the footprint of the pixel, sent to the ISS, HERA monitors (units containing single meaning only digital information is transferred out of Timepix detectors being developed at NASA for use on the pixels. Timepix can operate in event counting mode future MPCV missions), the proposed Biosentinel astro- (the base functionality of Medipix2, where a counter is biology deep-space cubesat mission and on trans-lunar incremented each time that during the shutter time period NASA-ORION missions in the 2020s. the charge deposited is over the designated threshold), The paper is organised as follows. In section 2 we arrival time mode (where essentially the rst time during describe the LUCID payload. In Section 3 we describe the shutter time period that the charge over a threshold the data reduction process, using GridPP resources, is deposited is recorded) and energy sensitive Time over processing all the data that was produced by the payload Threshold (ToT) mode (measuring the signal amplitude with a new machine learning analysis framework. We providing the per-pixel deposited energy). LUCID was also describe the development of the Timepix Analysis always run in ToT mode. Platform at School (TAPAS), a web platform to present the reduced LUCID data to secondary school students. In Section 4 we give the preliminary results that we have 2.2. LUCID Instrument Design gathered based on this data. In Section 5 we discuss our We brie y summarise the instrument design here; a results and future planned work. Finally, in Section 6, we full technical overview and detailed design information of summarise our ndings. LUCID is in the LUCID System Design Document (D. Cooke, SSTL, private communication). The payload has ve Timepix radiation detectors in a cube-like con gu- ration (see Fig. 1b), with four detectors orthogonally 2. The LUCID Payload positioned facing outwards (TPX0 through TPX3), and 2.1. LUCID and TechDemoSat-1 the fth in the centre (TPX4), facing outwards (relative to the centre of LUCID). A photograph of the instrument LUCID is a payload on the technology demonstra- is shown in Figure 1. The chips were surrounded by tion satellite TechDemoSat-1 (TDS-1, see gure 1). a 0.75mm thick aluminium dome which blocks intense The project started in 2008, and was developed as a light, plasma and low-energy charged particles. LUCID is collaboration between Langton Star Centre secondary mounted on the `Earthside' of TDS-1. school student researchers, the Medipix Collaboration, and Surrey Satellite Technology Limited (SSTL), who The detectors were calibrated by the Institute of built both LUCID and TDS-1. LUCID is part of the Experimental and Applied Physics (IEAP) at the Czech TDS-1 Space Environment Suite, which consists of the Technical University in Prague. The calibration process Miniature Radiation Environment and e ects Monitor involves exposing the detectors to 'X rays of discrete (MuREM, Taylor et al., 2012, Underwood et al., 2016), energy, and modelling the low energy end non-linear the Charged Particle Spectrometer (ChaPS, Kataria response of each individual pixel, see Jakubek et al. et al., 2013) and the Highly Miniaturized Radiation (2008) and Jakubek (2011). Monitor (HMRM, Mitchell et al., 2014, Guerrini et al., 2013). TDS-1 launched on 8 July 2014 (15:58:28 UTC) on The performance and expected measurements of a Soyuz-2-1b launch vehicle with Fregat-M upper stage LUCID were simulated in Whyntie and Harrison (2014), from the Baikonur Cosmodrome in Kazakhstan, into a Whyntie and Harrison (2015), who simulated the en- 635 km, 98.4 Sun-synchronous orbit. LUCID began vironment of LUCID and expected data rates using data collection shortly after launch, and data collection European Space Agency's (ESA) SPace ENVironment ceased on the 4th July 2017. TDS-1 operations have now Information System (SPENVIS, Heynderickx et al., 2004) ended, and at some point in the medium-term it will be and GEANT4 (Agostinelli et al., 2003, Allison et al., 2006 deorbited by the Icarus-1 Cran eld Drag Augmentation and Allison et al., 2016). System de-orbiter (Hobbs et al., 2013) which will over the next 25 years guide the spacecraft into the Earth's 2.3. LUCID Operations atmosphere, where it will disintegrate. Although the payload has ve detectors, for the The detectors used in LUCID are based on the Timepix majority of its time in orbit, only three were taking data ASIC chip (Llopart et al., 2007, Plackett et al., 2010), part at any one time. This was because LUCID would be of the second generation of Medipix (Medipix2, Llopart drawing too much power with all ve detectors taking 3 Figure 1: LUCID shown integrated on TDS-1 in the SSTL assembly room b) the LUCID payload showing the Timepix detector array (taken from the LUCID System Design Document). The four orthogonally positioned detectors are visible on the right hand side of the instrument. Table 1: Payload Speci cations Detectors 5 Timepix hybrid pixel detectors (300m silicon), 1 RADFET (metal-oxide-silicon) Power 1 permanent 5V supply and 1 28V supply for the Timepix detectors. With all 5 detectors running, maximum power usage is 8W Storage 2GB NAND ash Data transfer Maximum 20Mbit/s via X-Band Physical Dimensions 220mm x 135mm x 33mm (SSTL nano tray) Mass 1.2kg Dome 0.75mm aluminium, energies excluded; E > 0:4MeV and E > 10:0MeV (Whyntie and e p Harrison, 2015) 4 The nal data product from LUCID consists of 256x256 Table 2: Timepix con guration DAC-parameters, used for a large PNG images of each frame, CSV les that include the par- number of runs on all detectors (ID 321) ticle counts on a per frame basis, x,y,C formatted les for Active detectors 0, 1, 2, 3, 4 each frame, together with associated metadata and navi- Frame rate 1 per second gation and time stamp. This is post-conversion, as data Shutter Exposure Time 0.3 s was retrieved in a proprietary compressed raw format. Clock 33 MHz Number of frames 100 Bias Voltage 20.02 V 3. LUCID Data Analysis IKrum 1 Disc 127 Di erent particles typically produce very di erent Preamp 255 tracks when they pass through a Medipix detector (e.g. DAC Code 6 Llopart et al., 2002 and Bouchami et al., 2011). Given Sense DAC 0 the large volume of data collected by the instrument, it is Ext. DAC Sel. 0 necessary to automate the detection of di erent particles. Bu AnalogA 127 In this section we present two new high-performing Bu AnalogB 127 machine learning algorithms for classifying tracks in Hist 0 LUCID frames into particle types. This is a dicult THL Fine 300 task both because of the large quantity of data (meaning THL Coarse 7 that the algorithm has to be reasonably fast, and the VCAS 130 task cannot be done solely by human classi ers), as FBK 128 well as the general diculty in classifying tracks (essen- GND 80 tially a pattern recognition problem e.g. Holy et al., 2008). CTPR 0 THS 100 Firstly a clustering algorithm is run on the LUCID BiasLVDS 128 frames to identify individual particle tracks. After having RefLVDS 128 identi ed the track for each individual particle, we consid- ered two main methods classifying particles: data. Towards the end of operations all ve detectors were `Metric-Based Network' (MBN): we extract a given switched on and taking readings as the other payloads set of features for each particle track and then clas- entered di erent phases of operation. The detectors are sify the particle based on these metrics using ma- built and con gured to operate in sync - i.e. they capture chine learning. This network does not take per-pixel frames simultaneously. Speci cations are shown in table 1. energy distribution into account. `Deep learning': to use convolutional neural net- Each detector is able to independently capture a works to classify the particles tracks directly with `frame' of the radiation that passes through the detector a pure deep learning classi cation algorithm. This over some time period. LUCID was run with a range of network does take per-pixel energy distribution into frame rates and shutter exposure times, with the most account. commonly used shown in table 2, and a frame taken every second. The payload has a 2GB NAND ash storage Both of these methods are supervised machine learn- which is used to store data between passes of the ground ing algorithms and therefore the main limiting factor for stations, allowing LUCID to take data for more extended the development of the classi cation algorithms is the periods of time. The amount of available storage space computing power available and the amount of labelled and data transfer requirements of other experiments data that had been collected. In this work the training dictated frame rate. data is labelled by human classi ers. LUCID operated in a data gathering capacity from We also compare the performance of our algorithms to late October 2014 until July 2017. However, the rst an analytic classi er (i.e. tracks are classi ed based on an few months of operation were dedicated to instrument analytic function of the feature metrics) used in an early commissioning, and were used to nd the optimal con g- stage of the analysis of the data (see Whyntie et al., 2015). uration settings for data recording and the abilities of the payload. Nominal operations commenced in April 2015. Our approach can be compared to other studies of During this time, over 2.1 million frames were captured, pattern recognition and cluster analysis in Timepix de- over 82 runs and 11,700 les. A run is an 8 day capture tectors, such as Vilalta et al. (2011), Hoang et al. (2012), period, de ned by the operational schedule of TDS-1. Opalka et al. (2013) and Holy et al. (2008). Other possible machine learning approaches to classifying particles not 5 considered here include generative adversarial networks This allowed the latitude and longitude of LUCID when as well as autoencoders. the capture took place to be calculated. The classi cation was then run on all frames in the particular le, and for each frame alpha, beta, X-ray, proton, muon and other 3.1. Technical Implementation particle classi cations, and the latitude, longitude, frame number, capture timestamp were submitted directly to the The payload was operated by submitting Payload TAPAS (section 3.2) database via a POST request to a Task Request (PTR) les to SSTL. These les con- REST API endpoint. Each run and le (with its times- tain information about the start time of the run, the tamp, ID and con guration le used) was also submitted con guration le to use, and the overall schedule for this way. capturing data in this run. The LUCID con guration les specify which detectors should be used in this 3.2. The TAPAS Data Analysis and Visualisation Tool capture and the settings of each detector to be used, for instance the shutter exposure time and threshold value. We have developed The Timepix Analysis Platform at School (TAPAS) to allow secondary school student During the operating lifetime of the payload, data researchers across the UK to analyse and share the data was transferred from the satellite to SSTL and then that they gather using Timepix radiation detectors across downloaded to an IRIS server at the Langton Star all CERN@school projects, and additionally as a home Centre in Canterbury, UK. A cron job (a system used for for the particle count data from the LUCID experiment. scheduling jobs on UNIX based systems) ran every day The platform allows users to upload their own data, taken to check for the existence of new les on the SSTL FTP with the Timepix detectors using a software package server, and if les were detected, they were downloaded called Pixelman (Turecek et al., 2011a), or data which and processed immediately, converting them from a has been provided to them, such as the TimPix ISS-REM proprietary format for LUCID into individual frames of radiation data. A web application was chosen instead of x, y, C les (x-coordinate, y-coordinate, ToT value; tab a desktop application because it is very easy to access - separated). Metadata (time of capture, detector frame students only need a web browser and internet connection was from, the le the frame belongs to and location to use the service. information) was stored in an SQLite database. Once a user has uploaded a dataset, a cron job is used Pre-processed data was compressed and uploaded to run the analysis service every 10 minutes to process to GridPP (GridPP et al., 2006, Britton et al., 2009) the data. The analysis service uses multiple processes to storage, using the CernVM (Buncic et al., 2010). At the analyse large amounts of the data in a parallel fashion, same time as the uploading process, custom software was using the lucid utils LUCID algorithm. The service also developed to process runs in parallel, allowing many runs generates an image of every frame processed. TAPAS has to be analysed at the same time, exploiting the highly also been used to analyse small amounts of the LUCID parallel nature of running jobs on GridPP. data (the whole LUCID data set required the resources of GridPP). The processing software has been developed in Python, using Tensor ow (Abadi et al., 2016) machine learning for TAPAS is a Django framework based web application, particle analysis (see sections 3.5 and 3.6). The CernVM primarily written in Python for the page generation and File System (CVMFS) was used to deploy the software HTML, CSS and JavaScript for the frontend user facing dependencies and Python interpreter to the grid worker elements. The web platform has a MariaDB database nodes to run the software, and jobs were submitted backend that is used to store all metadata relating to a using the gLite middleware to a speci c GPU-backed job LUCID run, le and frame and particle classi cations. queue. The full source for the software is available at: This database also includes information and analysis https://github.com/willfurnell/lucid-grid/. results relating to user-made uploads. Each job running on a worker node copied data from The platform includes an API, using the `Django Rest a storage element to its working directory, extracted it Framework' to allow uploads and particle classi cations and then ran the classi cation algorithm. This process to be submitted using a programmatic REST (Represen- included getting frames and their metadata (namely the tational state transfer) interface. This is of particular capture time) from the pre-process database, and then cal- importance to the LUCID data, as this is how particle culating which of the Two Line Element les (TLE, les classi cations were submitted to the database from the that contain the information to track the location of the satellite ) had a date closest to the frame's timestamp. lines of data which can be used together with NORAD's SGP4/SDP4 orbital model to determine the position and velocity of the associated satellite.' - https://celestrak.com/columns/v04n03/ `A NORAD two-line element set consists of two 69-character 6 (typically IRIS secondary school researchers) to simply click through automatically generated questionnaires. The responses would be stored in a database alongside the pixel data and the metadata for nding out which frame the cluster belonged to. The LUCID Trainer web application is written in PHP for dynamic page generation and HTML, CSS and JavaScript for the frontend user interface. All the LUCID data is accessed via a custom REST API connecting the LUCID metadata database to the web application. The user response data is sent via an AJAX request to another PHP le that then stores the response in a simple SQLite database which was used due to its lightweight requirements. The training set used consisted of 1800 particles Figure 2: An example LUCID frame with individual particle tracks (tracks identi ed as in section 3.3). These tracks were identi ed. each classi ed once by a student researcher, with 24 classi ed as alpha particles, 988 as beta/electrons, 547 software running on GridPP worker nodes. as X-ray/photons, 27 as muons, 160 as protons and 54 as `other'. Clusters that the user could not identify, or We have programmed TAPAS to allow students to overlapping clusters were those classi ed as `other'. The download CSV les with particle count data, allow- student researcher classi ers themselves were trained ing them to conduct further analysis using their own using example tracks with known particle type, and choice of software packages - or even by choosing to reference diagrams similar to Figure 1 in Bouchami et al. write their own, in a programming language they are (2011) and Figure 1 in Whyntie et al. (2015). The student most comfortable with using. All of the LUCID par- researchers, and therefore the algorithm, may have had ticle count data is downloadable as CSV les from TAPAS. issues separating one particle type from another. The resulting training set is thus based on human classi ca- tions, which as a methodology is necessarily less accurate 3.3. Event Identi cation than using a training set constructed from known cali- brated sources. Nevertheless, these preliminary initial The rst step of classifying particles in a given Timepix results still give us a good overview of the distribution detector frame is to identify each individual particle track. of morphology of detections, and in the future calibrated The frame is fed in as a 256 x 256 matrix of energy values. classi cations can be generated and used for training to Non-zero pixels are systematically selected, and any obtain more realistic particle counts. non-zero pixels in the `ring' of the 8 adjacent (including diagonals) pixels are then added to the associated cluster. The nature of supervised machine learning algorithms For each adjacent pixel, all other adjacent pixels are trained on human classi cations means that the algorithm checked until there are no more pixels with non-zero can at best reproduce the classi cation that a human energy values. The clustering algorithm works through classi er would give. Similar projects such as Galaxy Zoo the entire frame and returns a list of clusters, each of (e.g. Lintott et al., 2008, Willett et al., 2017, Smethurst which is a list of pixels. An alternative clustering method et al., 2016) and the rest of the Zooniverse have shown is discussed in Granja et al. (2018), which uses track success in the classi cation of large samples of image shape and deposited energy values to classify the particle and image-like data based on a training set of human type, energy, and direction of events. Figure 2 shows an example LUCID frame where each particle track has been classi cations. Vilalta et al. (2011), in contrast to this successfully identi ed using the clustering algorithm. work, trained their algorithm using data taken at the Heavy Ion Medical ACcelerator Facility (HIMAC) in Chiba, Japan e.g. they red beams of known particle 3.4. Training Data type and energy at Medipix detectors, so they had a set of tracks labeled by the true particle properties (see also Both algorithms considered here are supervised, and Jakubek et al., 2010). Our human labelling of tracks will require that a subset of the data is labelled with the `true' be imperfect - however human labelling is still a valuable classi cation. To generate training data for classifying approach to develop for the classi cation of tracks as particle tracks, a web application called LUCID Trainer human classi ers can identify `unexpected' tracks that was created (Figure 3). It allows volunteer classi ers might be in the LUCID data but not in a laboratory 7 Figure 3: Screenshot from volunteer classi cation questionnaire, as seen by the student researchers. produced training set (e.g. Beck et al., 2018 found that a  Density - This is calculated using the calculate den- combination of human classi cation and machine learning sity function. classi cation gave the best results). Line Residual - This is calculated using the calculate non-linearity function. 3.5. Metric Based Network Curvature Radius and Circle Residual - These are The Metric Based Network (MBN) approach to calculated using the nd best t circle function. classifying the particles was to calculate a small number Average Neighbours - This is calculated using the of easily computed features for each track, and to then nd average neighbours function. classify the particles based on the metrics that had been calculated. Width - Number of Pixels / Diameter (if number of pixels > 1 else width = 0) The primary classi er used was the multi-layer neural network. The hyperparameters (number of hidden layers, The precise de nitions of these features are given in number of nodes in each hidden layer etc.) for this Appendix B. In early analyses of the data these metrics architecture was optimised manually. Other machine were used for an analytic classi cation of the tracks (e.g. learning classi ers were used for comparison as well as for if number of pixels less than 4, then classify as a X-ray ensuring that the limiting factor for the accuracy was the etc.), also described in Appendix B. size of the training set and not the architecture. The nal multi-layer neural network architecture has The general principles behind the multi-layer neural 8 inputs in the input layer (for the 8 metrics), 128 nodes network are described in Haykin (1998), and see the in the rst hidden layer with the sigmoid activation extensive literature on similar problem of classifying function, 48 nodes in the second hidden layer with the handwritten digits (a similar problem) e.g. McDonnell sigmoid activation function, and 6 outputs in the output et al. (2015). See Denby (1988) and Peterson (1989) layer with the softmax activation function, see gure 4. for early uses of neural networks in studying particle No dropout or regularisation was needed as the size of tracks in high energy density physics experiments, and the training dataset was already limited. The Gradient Farrell et al. (2017) for a more recent example in the LHC. Descent Optimiser was used as the training algorithm. The network was trained with a batch size (the number The eight metrics calculated from the pixel cluster rep- of training examples in one forward/backward pass) of resenting the particle track for the algorithm are: 128 particles. Additional classes can be added to the Metric-Based Network with ease as the outputs are Number of Pixels - This is calculated as the length one-hot encoded however it will have to be trained from of the pixels list. scratch on the dataset. Radius - This is calculated using the calculate radius function. 8 Figure 4: The architecture of the SML neural network used (diagram not to scale) 3.6. Deep Learning Deep learning algorithms present the possibility to Figure 5: Overall accuracy of each algorithm classify the particle tracks directly from the images (which contain more information) rather than from the tree and random forest, all available in Tensor ow) were pre-identi ed features considered in the MBN algorithm. also tested to con rm that our network architecture was Deep learning approaches have high requirements of optimal but are not shown in the gure. The hyperpa- computing power and typically are used with GPUs, so rameters for these algorithms were not investigated and we only present some early results here, although the the default parameters were used. approach could lead to much higher levels of accuracy with future work. Figure 5 shows the percentage correctly classi ed for a) the Metric-Based Network (MBN), b) a Deep Convolu- The deep learning architecture considered was a Deep tional Network (DCN) and c) the Analytic Classi er (AC). Convolutional Neural Network (CNN e.g. see Rawat and Wang, 2017 for a recent review) run for 5000 epochs. Epochs are one forward pass and one backward pass of all All performed between 80-95% accuracy, with the the training examples. We do not describe the algorithm MBN performing best on the test data. The AC had in great depth as the method remains preliminary. A a high accuracy of 84% but still less accurate than the Deep Residual Network (DRN e.g. He et al., 2016) has MBN. This is because the dataset consists of much also been tested on the data, but as of yet only for a small more beta and X-ray particles and therefore the overall number of epochs so has not yet attained a comparable accuracy appears to be high while its performance on a particle-by-particle basis was much poorer (see Fig high level of precision, but may be competitive in the 6b). Figure 6 shows the confusion matrices for our best future. algorithm, the Metric-Based Network, and the Analytic Classi er to show how machine learning leads to much Other advantages of the Deep Convolutional Network better performance. Each square shows the probability to the Metric-Based Network is that the Deep Convolu- one type of particle has of being classi ed as another. tional Network allows the energy values to be considered Complete success in classi cation would correspond to a (the feature metrics use for the MBN algorithm don't diagonal. The Metric-Based Network algorithm performs use the energy value of the pixels, just whether the pixel well, with the only substantial misclassi cations being was non-zero or not) and the convolutional lters allow 30-40% of muons and protons being misclassi ed as humans to visualise what the network has learned. electrons - partially because of similarities in the track shapes, partially because electrons dominate the overall sample, and partially because the labels used in the 3.7. Algorithm Performance training are imperfect. The Analytic Algorithm performs Figures 5 and 6 show the performance of the al- much poorer, misclassifying all particles as either electrons gorithms, the Metric Based Network (MBN) Deep and protons a substantial fraction of the time, and failing Convolutional Network (DCN) and the Analytic Classi er to identify any muons or `Others'. (AC). Other `o the shelf ' machine learning classi ers (support vector machines, k-nearest neighbours, a decision 9 Figure 6: Confusion matrices for the analytic classi er and MBN. Squares in the grid are colour coded by percentage of `actual class' classi ed as `predicted class'. 60 W) and the poles (roughly 60 N, 60 S) are clearly Table 3: Event Classi cations (all frames) evident for all particle types. We also show in gure 8 the Alpha 75902 ratio of heavy charged to light charged particles - this is Beta 169391026 non-uniform, showing that that heavy charged particles X-ray 369816217 have disproportionate intensity in the SAA and poles Proton 1065545 compared with the light charged particles. Muon 569771 Other 6703109 The aim in this section is to illustrate that LUCID is giving sensible results, as opposed to give detailed measurements of dose, linear energy transfer (LET) 4. Results and particle energy spectra. Although LUCID can In this section we present some preliminary results from make estimates of the energies of particles because the the data and particle classi cations described in Section 3. detectors have been calibrated, this analysis is beyond We use the classi cations from the Metric-Based Network the scope of this study and will be the focus of future work. algorithm e.g. the particles are classi ed only based on track morphology, and some tracks will be in incorrectly To create ux maps, a speci c date range was cho- classi ed, both due to the algorithm not having 100% ac- sen that covered the whole Earth, while also having a curacy (see Fig 6b) and imperfect labelling in the training homogeneous set of con guration les (also limiting the set. e ects of any time evolution of the radiation eld). The sub-set of the data considered consisted of 404 les, from 4.1. Classi cations 2016-08-17 to 2016-09-21, where we took the rst 10 frames from each of the les to result in a total of 4040 Table 3 presents the particle classi cations that we frames. For each le, the measured ux (count rate per have obtained for the whole lifetime of the LUCID pay- unit time per unit collecting area) was calculated. Then load. Di erent con guration les may produce di erent a K-dimensional (K-D) tree (with a maximum number classi cations due to saturating the frame (or conversely of neighbour lookups of 100) was used, so that for every having completely empty frames), and therefore these re- given latitude and longitude reached by LUCID, the sults should be taken as a preliminary analysis, rather than plotted ux at that point is the average of neighbouring con rmed results. Full payload data results are available frames ux, weighted inversely by the distance to the on TAPAS https://tapas.researchinschools.org frames. 4.2. Radiation Map and the South Atlantic Anomaly We plot in gure 7 the number of particles detected 4.3. Heavy Charged Particles over the Earth's surface for 4000 frames for di erent Of our tracks, about 3% were classi ed as `Other', and particles. Higher radiation levels (by more than a factor a substantial proportion of these particles are likely nuclei of ten) around South Atlantic Anomaly (an area of heavier than helium. We present example tracks in gures known increased radiation ux, centred at roughly 30 S, 10 Figure 7: Radiation maps for alpha particles, electrons, X-Rays, protons, muons and unclassi ed particles 11 Figure 8: A map of the ratio (Betas + Muons+1) / (Protons + Alphas+1). The +1 on the denominator to prevent division by 0 and +1 on numerator for a 1:1 ratio when both are 0. 12 9a and 9b, showing the extended shape. Timepix are par- 5.1. Other Experiments ticularly well suited for identifying heavy charged particles Our electron/proton percentages appear to be similar as they can extract track shape (as opposed to just particle to other experiments. Our radiation maps are similar to energy), see Granja et al. (2011), Stoe et al. (2012) and those found by the ISS-REMs (e.g. Fig 7. Stoe et al., Hoang et al. (2012). Identifying ux levels of heavy ions 2015). We do however go to higher latitudes (the ISS data carries important astrophysical information e.g. Aguilar only goes up to 55 N/S) mapping out the polar regions. et al. (2018). SATRAM, Granja et al. (2016) also map to the higher alti- tudes. Litvak et al. (2017) nd similar plots in the neutron 4.4. Preliminary Tests ux (measurements from the BTN-Neutron space experi- We do some early geometric tests to con rm that ment on the ISS), one of the particles LUCID can't detect. the experiment is gives reasonable results, as a prelude to future work investigating particle isotropy and the The ability of Timepix to classify heavy charged par- physics of particle transport from the sun and the trapped ticles is highlighted in Kroupa et al. (2015). Our results electron model. add to their ndings that Timepix have potential for use as a `heavy ion' monitor, or even telescope (Branchesi, Rather than travelling in straight lines, the particles 2016). Furthermore, our use of human classi ers provides are de ected towards the poles and the South Atlantic a testbed for future applications where humans are used to Anomaly (SAA) by the earth's magnetic eld, which is classify tracks that a traditional computer algorithm may why the particle count is higher as LUCID gets towards the struggle with. poles. We found slight evidence for marginally higher mea- sured number counts when LUCID is in front of the earth 5.2. Future Plans than behind the earth (relative to the Sun), mean of 7.1 Future work will focus on developing the algorithms particles per frame `dayside', and 6.3 particles `nightside'. and analyses presented in this work. In this paper we Dayside and nightside were de ned by calculating the an- focussed on the classi cation of track morphologies. A gle made by the LUCID satellite, the centre of the earth subsequent paper will explore the calibration of energy and the sun using the latitude, longitude and timestamp of measurements, and make dose maps, measurements of each frame. When this angle was between 0 and 90 , the LET, and particle energy spectra. We will also redo our frame was classi ed as on the `dayside' (i.e. the half of the analysis with larger training sets and improved imple- earth facing the sun) and when it was more than 90 it was mentations of the deep learning approaches. This new classi ed as on the `nightside'. Future work will focus on analysis will use particle tracks that have been validated investigating if this is a systematic in the data or not, and and calibrated from a de ned radiation eld so we can testing our results within current models of the Earth's identify events more accurately. magnetosphere, as the Earth's magnetosphere is known to be di erent on the `day' and `night' sides of the Earth. We will investigate di erences in ux measured in Pressure from the solar wind coming from the sun on the the di erent detectors (e.g. di erences between Timepix dayside compresses the magnetic eld, and on the night- perpendicular to each other). In addition, there is still side elongates it. In the Van Allen radiation belts (which much to exploit from the ability of Medipix detectors LUCID is close to/passes through in the SAA), charged to measure both the elevation and azimuthal angle of particles are trapped within the magnetic eld, and where particle trajectories in space e.g George (2014), Kroupa this eld is condensed on the dayside there will be a higher et al. (2015) nd that particles are highly directional particle density, which is consistent with a higher parti- in the SAA (c.f. Granja et al., 2014b) See also rocket cle count on this side of the earth (e.g. Williams and experiments at lower altitudes e.g. Z abori et al. (2016). Mead, 1965, Domrachev and Chugunin, 2002, Khazanov This would require identi cation of particle track angle and Liemohn, 2002). These results show that the LUCID and inclination as part of the machine learning classi- data can contribute to testing and update models of space cation algorithm, although we note that this must be weather. done carefully to avoid uncertainty regarding the angles of entry. In addition it would also be worthwhile to compare the student researcher track labels to labelled 5. Discussion tracks from known sources - as already mentioned, the In this section we discuss how our results compare to classi cations of Vilalta et al. (2011) and Hoang (2013) comparable space-based space radiation detectors, what have the advantage of `true' labels compared with our our results mean for future use of Medipix in space, and human-classi ed labels, but may nd it more dicult to what our future plans with LUCID data are. identify unexpected tracks. With classi cations of particle inclination as well as particle type we will use the particle angle measurements 13 Figure 9: Sample tracks of likely heavy charged particles. The colour scale is linear and scales with the energy deposited by the ion in the silicon layer for each pixel. to test the isotropy of particle ux in the trapped electron ent types of radiation and the inverse square law). IRIS is model. We will also investigate evolution in ux over currently undertaking pedagogical research about the ef- time, to link measurements to solar cycles (Thomas et al., fectiveness of this approach to teaching, and is developing 2014), and investigate the growth of the SAA and any plans to expand CERN@school to other CERN member possible link to geomagnetic reversal (Pav on-Carrasco states. and De Santis, 2016). 6. Conclusions It has recently been shown that the kinetic energy of particles can be reconstructed using a Timepix detector In this paper we introduce the LUCID detector (the (Kroupa et al., 2018), which will also be a topic of future third use of Medipix detectors in space, the second in open investigation with the information from the data that we space, and the rst on a commercial platform and with have obtained from LUCID. a 3D con guration) that ew aboard TDS-1 taking data from 2014 to 2017. We describe the data pipeline from 5.3. Lessons From CERN@school data collection to reduced catalogue of classi ed particles, We view LUCID as technological test of Medipix a novel machine learning particle track classi cation detectors for future space missions, giving Timepix data algorithm, and some early science results. We also discuss at a di erent altitude to other experiments, for the rst LUCIDs links to the larger CERN@school ecosystem. time on a commercial platform, in a novel con guration of chips. However we also view CERN@school as an The payload was operated by submitting Payload Task important test of the use of Medipix technology. Medipix Requests to SSTL. Data immediately after collection was had previously been predominantly used by professional stored on a ash partition on the spacecraft until passes scientists. CERN@school was the rst widespread use of over ground stations. The data was passed from the Medipix across 100s of institutions, with users ranging ground station to SSTL, and then to IRIS servers. From from novices to experts. The challenges of handling data there the LUCID frame data were passed to GridPP for from a very heterogeneous set of sources for use by a processing, and the reduced data passed back to the IRIS wide variety of users of di erent levels of expertise is servers. a test bed for any future hypothetical large-scale use of Medipix outside of academia and research labs e.g. We use machine learning techniques to classify Medipix as a personal radiation monitor in a nuclear power Timepix particle tracks, presenting two types of al- plant or for nuclear medicine workers (Michel et al., 2009). gorithm: a) a metric-based neural network classi er that uses eight pre-identi ed `features' and b) a deep Alongside the research applications, CERN@school de- learning approach, showing both perform well. These tectors in schools also have an educational role (e.g. as a algorithms perform competitively compared with exist- replacement for a geiger counter in teaching about di er- ing analytic algorithms used on large sets of Medipix data. 14 for calibrating the Timepix detectors used in LUCID. We present TAPAS, a new platform for Medipix/Timepix data that has been used extensively by secondary school Alongside the authors, the following students have also researchers in the UK for managing large amounts of contributed substantially over the history of the project: radiological data from a wide range of sources. TAPAS Toby Freeland, Matt Harrison, Cal Hewitt, Sam Kittle, permits easy storage, sharing and visualisation via a Nick Liu, Rachel O'Leary, Rachel Powell, Adam Sandey, browser interface. Hector Stalker, Tom Stevenson and Cassie Warren, as well as extensive support and encouragement over many Finally we discuss some early results from the reduced years from Dr Tom Whyntie (Oxford). Thank you also LUCID data. We nd that the majority of particles to students who contributed to the classi cations used in detected are electrons, in agreement with the trapped the training set for the machine learning classi cations, electron model and earlier simulations of LUCID. Future and the many other students who worked on LUCID over work will focus on using LUCID data to characterise the last ten years. the radiation environment of LEO (in particular using angular information), and contribute towards planned A huge personal thank you from the authors to Prof. future use of Medipix on space missions. Becky Parker (IRIS) and Laura Thomas (IRIS) who have made LUCID and CERN@school happen, and have Key results: inspired countless students across the country over many years. We present the rst results from LUCID, a novel space radiation detector that use Timepix detectors AS and EF would like to thank Mr Rupert Champion in orbit on TechDemoSat-1 (Langton Star Centre) and Dr Tim Lesworth (Langton Star Centre) for providing ongoing support with their We have developed the TAPAS platform, a versa- IRIS activities. PH acknowledges funding from the tile tool for managing Timepix datasets from a wide Engineering and Physical Sciences Research Council. range of sources. We used TAPAS to handle large amounts of LUCID and CERN@school data, lay- The authors would like to thank Prof. Steven Rose ing the groundwork for more widespread adoption (Imperial/Oxford) and Dr Carlos Granja (Advacam, of Timepix Prague) for their helpful comments, remarks and input to We present the use of machine learning algorithms the paper. to classify track morphology and thus particle type in Medipix data, allowing for quick, accurate particle The authors and the Institute for Research in classi cation for huge amounts of data Schools would like to acknowledge generous support from: Humphrey Battcock, The Science and Technology Our results add to the evidence that Medipix Facilities Council, The Science Museum, The Institute detectors are well suited as space weather moni- of Physics, The Royal Commission for the Exhibi- tors/detectors tion of 1851, The Ogden Trust, CERN, The Medipix Collaboration, NASA, the UK Space Agency, Kent We have made preliminary ux maps (showing the County Council, IEAP CTU Prague, SEEDA and the SAA and similar features) at an altitude of 600km, GridPP Collaboration (in particular Dr Dan Traynor which appear similar to those of other detectors and and Queen Mary University of London for use of their present several other early science results, including GPU and storage resources, and Professor Steve Lloyd detecting heavier charged particles for supporting and enabling schools to work with GridPP). CERN@school has acted as a large scale trial of the use of Medipix in diverse environments by sizeable The authors would also like to gratefully thank the numbers of users anonymous referees whose comments greatly improved the quality of this paper. Acknowledgements The authors and IRIS are extremely grateful to the Appendix A. Code Access Medipix Collaboration and SSTL for ten years of support, in particular we are hugely thankful to Dr Michael Camp- All the code discussed in this work is open source bell (CERN/Medipix Collaboration), Prof. Larry Pinksy and available online: https://github.com/amshenoy/ (University of Houston/NASA), David Cooke (SSTL), lucid_neural_analysis Dr Stuart Eves (SSTL) and Dr Jonathan Eastwood https://github.com/willfurnell/lucid-grid/ (Imperial). Speci c thanks to Prof. Stanislav Pospisil https://github.com/InstituteForResearchInSchools/ and IEAP CTU Prague for their support, in particular 15 lucid_utils rather than around them. If the cluster is only one pixel in size, then the density is by default set to 1. Information about LUCID, some frames from the Calculate Non-Linearity (Line Residual) - This func- mission, and some visualisation tools can be found at: tion returns the angle  anticlockwise from the x- http://starserver.researchinschools.org/lucid_ axis, with the line passing through the cluster cen- dashboard troid. First, all the coordinates of the pixels within To support and assist the production of labelled LU- the cluster are split into separate lists of X and Y CID data, the training web application can be found at: values. Single pixel tracks are by default given an an- http://starserver.researchinschools.org/lucid_ gle and line residual of 0 as these are completely lin- trainer ear. For all other clusters, the above least squares re- The TAPAS platform can be found at https://tapas. gression function is used to calculate the line of best researchinschools.org/. IRIS can be contacted for in- t in the form . From the line of best t, the best- t quiries about accessing LUCID data via TAPAS at the angle (anti-clockwise from the x-axis) of the line-of- supplied email address. best- t can be calculated using simple trigonometry. XY XY b = (X) Appendix B. Feature De nitions This appendix gives the de nitions of the features used x a = Y bX; where X = and Y = n n in the `Metric-Based Network' algorithm in Section 3, and the details of the analytic classi er. P PixelCount Line Residual = (x sin  y cos n n n=0 x sin  + y cos ) c c Find Average Neighbours - For each individual pixel in the cluster, the number of neighbouring pixels is counted by iterating through the surrounding pixel This best t angle is used to calculate the line resid- coordinates and checking if they exist in the cluster. ual value using the equation above. The line residual The pixels that exist are then stored in a list as these value is calculated by the sum of the squares of the are the ones that have an energy deposit. The mean distance from each pixel coordinate to the line travel- of the items in the list is then calculated and the ling through the centroid where the line is calculated function returns the average number of neighbours. using the best t angle . The equation above shows the simplest form of the line residual where (x , y ) c c Find Centroid (Centroid) - The centroid of the clus- is the coordinate of the centroid and (x , y ) is the n n ter is the centre of the particle track. This centroid is coordinate of the pixels in the cluster where n is the found by calculating the mean of the x values and the index of the pixel in the cluster. mean of the y values of the pixels in the cluster. The mean x-value is the presumed to be the x-coordinate Find Best Fit Circle (Curvature Radius & Circle of the centre point and the mean y-value is presumed Residual) - This function is used to perform circle to be the y-coordinate of the centre point. This co- regression. For single pixel tracks, this cannot be ordinate is not the centre of the smallest enclosing done and therefore single pixels are by default given circle but in fact, it is the centre of mass with no values of zero. As the cluster is a very poor initial weighting on energy values. This function is used to guess for the centre of the circle, multiple test calculate an auxiliary metric as the centroid is used circles are generated using the calculated best t to generate other metrics but it is not explicitly used angle from the regression line calculation. For each for classi cation as it does not have any informative of these test circles, the program performs circle value by itself. regression on the cluster using least squares. The Euclidean distance between the data points and the Calculate Radius (Radius & Diameter) - mean circle centred at (x , y ) is calculated. The p c c 2 2 Radius = (x x ) + (y y ) c n c n mean distance is then subtracted from all of the calculated distances. This value is optimised using The Euclidean distance between the coordinates of the least squares function. The optimisation process each pixel to the centroid is calculated. The radius returns an optimised centre point. This optimised is set to the highest distance from the centroid. The centre point is used to calculate the distance of all diameter is then simply twice the calculated radius. the points from the centre of the circle once again. Calculate Density (Density) - The mean distance is set as the new radius and PixelCount the residual is calculated by the summation of the Density = squares of the di erence between the distance from The density can be greater than 1 as the cluster's each point to the optimised centre point and the radius passes through the centre of the outer pixels mean radius. 16 p 2 2 (x x ) +(y y ) c n c n (Mean Radius) R = PixelCount 2 2 Circle Residual =  ( (x x ) + (y y ) c n c n n=0 R ) The optimised test circles are then organised in order of the magnitude of the residual. The aim of the optimisation is to reduce the circle residual as much as possible and therefore the circle with the least residual is chosen as the best t circle. Analytic Classi er - The analytic classi er for parti- cles used in early LUCID analysis is based on metrics that can be found in the open source lucid utils library: https: //github.com/InstituteForResearchInSchools/ lucid_utils/blob/master/lucid_utils/classification/ lucid_algorithm.py 17 REFERENCES Medipix3 pixel readout chip. Journal of Instrumentation 6 (01), C01052{C01052. URL http://stacks.iop.org/1748-0221/6/i=01/a=C01052?key= crossref.7800cb0057377fce4b91848ea8e839f1 References Beck, M. R., Scarlata, C., Fortson, L. F., Lintott, C. J., Simmons, B. D., Galloway, M. A., et al., 6 2018. Integrating human and Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, machine intelligence in galaxy morphology classi cation tasks. C., et al., 3 2016. TensorFlow: Large-Scale Machine Learning on Monthly Notices of the Royal Astronomical Society 476 (4), Heterogeneous Distributed Systems. 5516{5534. URL http://arxiv.org/abs/1603.04467 URL https://academic.oup.com/mnras/article/476/4/5516/ Accorsi, R., Celentano, L., Laccetti, P., Lanza, R. C., Marotta, M., Mettivier, G., et al., 2008. High-Resolution $^f125g$I Small Ani- Bell, A. R., Araudo, A. T., Matthews, J. H., Blundell, K. M., 1 mal Imaging With a Coded Aperture and a Hybrid Pixel Detector. 2018. Cosmic-ray acceleration by relativistic shocks: limits and IEEE Transactions on Nuclear Science 55 (1), 481{490. estimates. Monthly Notices of the Royal Astronomical Society URL http://ieeexplore.ieee.org/document/4448434/ 473 (2), 2364{2371. Ackermann, M., Ajello, M., Allafort, A., Baldini, L., Ballet, J., Bar- URL https://academic.oup.com/mnras/article/473/2/2364/ biellini, G., et al., 2 2013. Detection of the characteristic pion- decay signature in supernova remnants. Science (New York, N.Y.) Bisogni, M. G., Campbell, M., Conti, M., Delogu, P., Fantacci, 339 (6121), 807{11. M. E., Heijne, E. H. M., et al., 11 1998. Performance of a 4096- URL http://www.ncbi.nlm.nih.gov/pubmed/23413352 pixel photon counting chip&lt;/title&gt;. Vol. 3445. International Agostinelli, S., Allison, J., Amako, K., Apostolakis, J., Araujo, Society for Optics and Photonics, pp. 298{304. H., Arce, P., et al., 7 2003. Geant4a simulation toolkit. Nuclear URL http://proceedings.spiedigitallibrary.org/ Instruments and Methods in Physics Research Section A: Ac- proceeding.aspx?articleid=957789 celerators, Spectrometers, Detectors and Associated Equipment Blanchot, G., Chmeissani, M., D az, A., D az, F., Fern andez, J., 506 (3), 250{303. Garc a, E., et al., 12 2006. Dear-Mama: A photon counting X-ray URL https://www.sciencedirect.com/science/article/pii/ imaging project for medical applications. Nuclear Instruments S0168900203013688 and Methods in Physics Research Section A: Accelerators, Aguilar, M., Alberti, G., Alpat, B., Alvino, A., Ambrosi, G., Spectrometers, Detectors and Associated Equipment 569 (1), Andeen, K., et al., 4 2013. First Result from the Alpha Magnetic 136{139. Spectrometer on the International Space Station: Precision URL https://www.sciencedirect.com/science/article/pii/ Measurement of the Positron Fraction in Primary Cosmic Rays S0168900206016147 of 0.5350 GeV. Physical Review Letters 110 (14), 141102. Blasi, P., 11 2013. The origin of galactic cosmic rays. The Astronomy URL https://link.aps.org/doi/10.1103/PhysRevLett.110. and Astrophysics Review 21 (1), 70. URL http://link.springer.com/10.1007/s00159-013-0070-7 Aguilar, M., Ali Cavasonza, L., Ambrosi, G., Arruda, L., Attig, Bouchami, J., Guti errez, A., Holy, T., Houdayer, A., Jakubek, J., N., Aupetit, S., et al., jan 2018. Observation of New Properties Lebel, C., et al., 5 2011. Measurement of pattern recognition of Secondary Cosmic Rays Lithium, Beryllium, and Boron by eciency of tracks generated by ionizing radiation in a Medipix2 the Alpha Magnetic Spectrometer on the International Space device. Nuclear Instruments and Methods in Physics Research Station. Physical Review Letters 120 (2), 021101. Section A: Accelerators, Spectrometers, Detectors and Associated URL https://link.aps.org/doi/10.1103/PhysRevLett.120. Equipment 633, S187{S189. URL https://www.sciencedirect.com/science/article/pii/ Allison, J., Amako, K., Apostolakis, J., Araujo, H., Arce Dubois, P., S0168900210013537 Asai, M., et al., 2 2006. Geant4 developments and applications. Branchesi, M., 5 2016. Multi-messenger astronomy: gravitational IEEE Transactions on Nuclear Science 53 (1), 270{278. waves, neutrinos, photons, and cosmic rays. Journal of Physics: URL http://ieeexplore.ieee.org/document/1610988/ Conference Series 718 (2), 022004. Allison, J., Amako, K., Apostolakis, J., Arce, P., Asai, M., Aso, URL http://stacks.iop.org/1742-6596/718/i=2/a=022004? T., et al., 11 2016. Recent developments in Geant4. Nuclear key=crossref.f4836af5970c7a2c85257dcd576d3ecc Instruments and Methods in Physics Research Section A: Ac- Britton, D., Cass, A., Clarke, P., Coles, J., Colling, D., Doyle, celerators, Spectrometers, Detectors and Associated Equipment A., et al., 6 2009. GridPP: the UK grid for particle physics. 835, 186{225. Philosophical Transactions of the Royal Society A: Mathematical, URL https://www.sciencedirect.com/science/article/pii/ Physical and Engineering Sciences 367 (1897), 2447{2457. S0168900216306957 URL http://rsta.royalsocietypublishing.org/cgi/doi/10. Amato, E., Blasi, P., 4 2017. Cosmic ray transport in the Galaxy: 1098/rsta.2009.0036 A review. Advances in Space Research. Buncic, P., Sanchez, C. A., Blomer, J., Franco, L., Harutyunian, A., URL https://www.sciencedirect.com/science/article/pii/ Mato, P., et al., 4 2010. CernVM a virtual software appliance for S0273117717302995 LHC applications. Journal of Physics: Conference Series 219 (4), Amendolia, S., Bertolucci, E., Bisogni, M., Bottigli, U., Ceccopieri, A., Ciocci, M., et al., 2 1999. MEDIPIX: a VLSI chip for a URL http://stacks.iop.org/1742-6596/219/i=4/a=042003? GaAs pixel detector for digital radiology. Nuclear Instruments key=crossref.07e37c1115c8c3c79a47061fd61faaf8 and Methods in Physics Research Section A: Accelerators, Butzer, J. S., Butler, A. P., Butler, P., Bones, P., Cook, N., Tlustos, Spectrometers, Detectors and Associated Equipment 422 (1-3), L., 11 2008. Medipix imaging - evaluation of datasets with PCA. 201{205. In: 2008 23rd International Conference Image and Vision Com- URL https://www.sciencedirect.com/science/article/pii/ puting New Zealand. IEEE, pp. 1{6. S0168900298010948 URL http://ieeexplore.ieee.org/document/4762080/ Baca, T., Platkevic, M., Jakubek, J., Inneman, A., Stehlikova, Campbell, M., Heijne, E., Meddeler, G., Pernigotti, E., Snoeys, W., V., Urban, M., et al., 10 2016. Miniaturized X-ray telescope for 6 1998. A readout chip for a 6464 pixel matrix with 15-bit single VZLUSAT-1 nanosatellite with Timepix detector. Journal of photon counting. IEEE Transactions on Nuclear Science 45 (3), Instrumentation 11 (10), C10007{C10007. 751{753. URL http://stacks.iop.org/1748-0221/11/i=10/a=C10007? URL http://ieeexplore.ieee.org/document/682629/ key=crossref.c40da9dc58b922546561e9a2917ae846 Castellina, A., Donato, F., 2013. Astrophysics of Galactic Charged Ballabriga, R., Blaj, G., Campbell, M., Fiederle, M., Grei enberg, Cosmic Rays. In: Planets, Stars and Stellar Systems. Springer D., Heijne, E. H. M., et al., 1 2011. Characterization of the 18 Netherlands, Dordrecht, pp. 725{788. Giuliani, A., Cardillo, M., Tavani, M., Fukui, Y., Yoshiike, S., Torii, URL http://link.springer.com/10.1007/978-94-007-5612-0_ K., et al., nov 2011. Neutral pion emission from accelerated pro- 14 tons in the supernova remnant W44. The Astrophysical Journal Choi, H.-S., Lee, J., Cho, K.-S., Kwak, Y.-S., Cho, I.-H., Park, Y.- Letters, Volume 742, Issue 2, article id. L30, 5 pp. (2011). 742. D., et al., 6 2011. Analysis of GEO spacecraft anomalies: Space URL http://arxiv.org/abs/1111.4868http://dx.doi.org/10. weather relationships. Space Weather 9 (6), n/a{n/a. 1088/2041-8205/742/2/L30 URL http://doi.wiley.com/10.1029/2010SW000597 Granja, C., Jakubek, J., K oster, U., Platkevic, M., Pospisil, S., Collins, P., Akiba, K., Alexander, M., Artuso, M., Bayer, F., van 5 2011. Response of the pixel detector Timepix to heavy ions. Beuzekom, M., et al., 4 2011. The LHCb VELO upgrade. Nuclear Nuclear Instruments and Methods in Physics Research Section Instruments and Methods in Physics Research Section A: Ac- A: Accelerators, Spectrometers, Detectors and Associated Equip- celerators, Spectrometers, Detectors and Associated Equipment ment 633, S198{S202. 636 (1), S185{S193. URL https://www.sciencedirect.com/science/article/pii/ URL https://www.sciencedirect.com/science/article/pii/ S0168900210013562 S0168900210009629 Granja, C., Jakubek, J., Polansky, S., Zach, V., Krist, P., Chvatil, Colthurst, D., Kent, M., Mauchaza, R., Senasinghe, T., Sparkes, D., et al., nov 2018. Resolving power of pixel detector Timepix G., Vasilijevic, F., et al., 2015. Research science in schools: An for wide-range electron, proton and ion detection. Nuclear Instru- insight into the student experience in biomedical science and par- ments and Methods in Physics Research Section A: Accelerators, ticle physics. School Science Review 97 (358), 4451. Spectrometers, Detectors and Associated Equipment 908, 60{71. Cucinotta, F. A., 1 2007. Space Radiation Organ Doses for Astro- URL https://www.sciencedirect.com/science/article/pii/ nauts on Past and Future Missions. S0168900218309586 URL https://ntrs.nasa.gov/search.jsp?R=20070010704 Granja, C., Platkevic, M., Turecek, D., Kraus, V., Pospisil, S., Cucinotta, F. A., Manuel, F. K., Jones, J., Iszard, G., Murrey, J., Jakubek, J., et al., 2014a. Timepix-Based Miniaturized Radiation Djojonegro, B., et al., 11 2001. Space radiation and cataracts in Micro-Tracker for the Micro-Satellite RISESAT. Transactions of astronauts. Radiation research 156 (5 Pt 1), 460{6. the Japan Society for Aeronautical and Space Sciences, Aerospace URL http://www.ncbi.nlm.nih.gov/pubmed/11604058 Technology Japan 12 (ists29), Tr 7{Tr 11. D aniel, V., P na, L., Inneman, A., Zadra zil, V., B a ca, T., Platkevi c, URL http://jlc.jst.go.jp/DN/JST.JSTAGE/tastj/12.Tr_7? M., et al., 9 2016. Terrestrial gamma-ray ashes monitor demon- lang=en&from=CrossRef&type=abstract strator on CubeSat. Vol. 9978. International Society for Optics Granja, C., Polansky, S., Pospisil, S., Vykydal, Z., Turecek, D., and Photonics, p. 99780D. Owens, A., et al., 2014b. Characterization and directional visu- URL http://proceedings.spiedigitallibrary.org/ alization of space radiation quanta in low earth orbit with the proceeding.aspx?doi=10.1117/12.2240299 compact spacecraft payload SATRAM. In: Proceedings of Sci- Deligny, O., O., dec 2016. Cosmic-Ray Anisotropies: A Review. ence. eprint arXiv:1612.08002. Granja, C., Polansky, S., Vykydal, Z., Pospisil, S., Owens, A., URL http://arxiv.org/abs/1612.08002 Kozacek, Z., et al., 6 2016. The SATRAM Timepix spacecraft Denby, B., 6 1988. Neural networks and cellular automata in exper- payload in open space on board the Proba-V satellite for wide imental high energy physics. Computer Physics Communications range radiation monitoring in LEO orbit. Planetary and Space 49 (3), 429{448. Science 125, 114{129. URL https://www.sciencedirect.com/science/article/pii/ URL https://www.sciencedirect.com/science/article/pii/ 0010465588900045 S0032063316300216 Domrachev, V. V., Chugunin, D. V., 2002. Modeling of Plasma Den- Grei enberg, D., Fiederle, M., Vykydal, Z., Kr al, V., Jakubek, J., sity in the Earth's Dayside Inner Magnetosphere. Cosmic Research Holy,  T., et al., 8 2009. Detection eciency of ATLAS-MPX 40 (4), 411{417. detectors with respect to neutrons. Nuclear Instruments and URL http://link.springer.com/10.1023/A:1019806814334 Methods in Physics Research Section A: Accelerators, Spectrom- Farrell, S., Anderson, D., Cala ura, P., Cerati, G., Gray, L., eters, Detectors and Associated Equipment 607 (1), 38{40. Kowalkowski, J., et al., 8 2017. The HEP.TrkX Project: deep URL https://www.sciencedirect.com/science/article/pii/ neural networks for HL-LHC online and oine tracking. EPJ S0168900209005968 Web of Conferences 150, 00003. Greisen, K., 4 1966. End to the Cosmic-Ray Spectrum? Physical URL http://www.epj-conferences.org/10.1051/epjconf/ Review Letters 16 (17), 748{750. 201715000003 URL https://link.aps.org/doi/10.1103/PhysRevLett.16.748 Ferrari, F., Szuszkiewicz, E., may 2009. Cosmic Rays: A Review for GridPP, T. G., Faulkner, P. J. W., Lowe, L. S., Tan, C. L. A., Astrobiologists. Astrobiology 9 (4), 413{436. Watkins, P. M., Bailey, D. S., et al., 1 2006. GridPP: development URL http://www.ncbi.nlm.nih.gov/pubmed/19519216http: of the UK computing Grid for particle physics. Journal of Physics //www.liebertonline.com/doi/abs/10.1089/ast.2007.0205 G: Nuclear and Particle Physics 32 (1), N1{N20. Frisch, D. H., Smith, J. H., 5 1963. Measurement of the Relativis- URL http://stacks.iop.org/0954-3899/32/i=1/a=N01?key= tic Time Dilation Using -Mesons. American Journal of Physics crossref.9896604cdb8b56fd9e2febdf74336b83 31 (5), 342{355. Guerrini, N., Turchetta, R., Grin, D., Morse, T., Morse, A., URL http://aapt.scitation.org/doi/10.1119/1.1969508 Poyntz-Wright, O., et al., 12 2013. Design and characterisation Gaza, R., Kroupa, M., Rios, R., Stoe, N., Benton, E. R., Semones, of a highly miniaturised radiation monitor HMRM. Nuclear E. J., 11 2017. Comparison of novel active semiconductor pixel Instruments and Methods in Physics Research Section A: Ac- detectorwith passive radiation detectors during the NASA Orion celerators, Spectrometers, Detectors and Associated Equipment ExplorationFlight Test 1 (EFT-1). Radiation Measurements 106, 731, 154{159. 290{297. URL https://www.sciencedirect.com/science/article/pii/ URL https://www.sciencedirect.com/science/article/pii/ S0168900213009066?via=ihub S135044871730224X#bib22 Hartmann, B., 2013. A Novel Approach to Ion Spectroscopy of George, S. P., 2014. Mapping the Asymmetry in the SAA Fluxes Therapeutic Ion Beams Using a Pixelated Semiconductor Detec- Using the Medipix Particle-by-Particle Directional Information. tor - heiDOK. Ph.D. thesis. In: 19th WRMISS Workshop. URL http://archiv.ub.uni-heidelberg.de/volltextserver/ Ginzburg, V. L., 2 1996. Cosmic ray astrophysics (history and 15690/ general review). Physics-Uspekhi 39 (2), 155{168. Haykin, S., 1998. Neural Networks: A Comprehensive Foundation, URL http://stacks.iop.org/1063-7869/39/i=2/a=A03?key= 2nd Edition. Prentice Hall PTR, Upper Saddle River, NJ, USA. crossref.c85a1188a6f128233207236dd01fb4d4 He, K., Zhang, X., Ren, S., Sun, J., 6 2016. Deep Residual Learning 19 for Image Recognition. In: 2016 IEEE Conference on Computer Energy Cosmic Rays. Vision and Pattern Recognition (CVPR). IEEE, pp. 770{778. URL http://arxiv.org/abs/1101.4256http://dx.doi.org/10. URL http://ieeexplore.ieee.org/document/7780459/ 1146/annurev-astro-081710-102620 Heynderickx, D., Quaghebeur, B., Wera, J., Daly, E. J., Evans, H. Kroupa, M., Bahadori, A., Campbell-Ricketts, T., Empl, A., Hoang, D. R., 10 2004. New radiation environment and e ects models S. M., Idarraga-Munoz, J., et al., 7 2015. A semiconductor in the European Space Agency's Space Environment Information radiation imaging pixel detector for space radiation dosimetry. System (SPENVIS). Space Weather 2 (10), n/a{n/a. Life Sciences in Space Research 6, 69{78. URL http://doi.wiley.com/10.1029/2004SW000073 URL https://www.sciencedirect.com/science/article/pii/ Hoang, S., Pinsky, L., Vilalta, R., Jakubek, J., 12 2012. LET S2214552415000620 Estimation of Heavy Ion Particles based on a Timepix-Based Si Kroupa, M., Bahadori, A. A., Campbell-Ricketts, T., George, S. P., Detector. Journal of Physics: Conference Series 396 (2), 022023. Zeitlin, C., mar 2018. Kinetic energy reconstruction with a single URL http://stacks.iop.org/1742-6596/396/i=2/a=022023? layer particle telescope. Applied Physics Letters 112 (13), 134103. key=crossref.74ffd9c63c4ad722b5ea2caa0c667142 URL http://aip.scitation.org/doi/10.1063/1.5024920 Hoang, S. M., 2013. A PATTERN RECOGNITION APPROACH Kuwahara, T., Yoshida, K., Sakamoto, Y., Tomioka, Y., Fukuda, TO LEARNING TRACKS OF HEAVY-ION PARTICLES IN K., 12 2011. Satellite system integration based on Space Plug and TIMEPIX DETECTORS. Ph.D. thesis, University of Houston. Play Avionics. In: 2011 IEEE/SICE International Symposium on URL https://uh-ir.tdl.org/uh-ir/handle/10657/1189 System Integration (SII). IEEE, pp. 896{901. Hobbs, S., Kingston, J., Roberts, P., Juanes, C., Sewell, R., Snapir, URL http://ieeexplore.ieee.org/document/6147568/ B., et al., 1 2013. De-Orbit Sail Design for TechDemoSat-1. Li, X., Palo, S., Kohnert, R., Blum, L., Gerhardt, D., Schiller, Q., Holy, T., Heijne, E., Jakubek, J., Pospisil, S., Uher, J., Vykydal, et al., feb 2013. Small Mission Accomplished by Students-Big Im- Z., jun 2008. Pattern recognition of tracks induced by individual pact on Space Weather Research. Space Weather 11 (2), 55{56. quanta of ionizing radiation in Medipix2 silicon detector. Nuclear URL http://doi.wiley.com/10.1002/swe.20025 Instruments and Methods in Physics Research Section A: Ac- Lintott, C. J., Schawinski, K., Slosar, A., Land, K., Bamford, S., celerators, Spectrometers, Detectors and Associated Equipment Thomas, D., et al., 2008. Galaxy Zoo: morphologies derived from 591 (1), 287{290. visual inspection of galaxies from the Sloan Digital Sky Survey. URL https://www.sciencedirect.com/science/article/pii/ Monthly Notices of the Royal Astronomical Society, Volume 389, S0168900208004592 Issue 3, pp. 1179-1189. 389, 1179{1189. Jakubek, J., 5 2011. Precise energy calibration of pixel detector URL http://adsabs.harvard.edu/abs/2008MNRAS.389.1179L working in time-over-threshold mode. Nuclear Instruments and Litvak, M. L., Mitrofanov, I. G., Nuzhdin, I. O., Vostrukhin, A. V., Methods in Physics Research Section A: Accelerators, Spectrom- Golovin, D. V., Kozyrev, A. S., et al., 3 2017. Monitoring of the eters, Detectors and Associated Equipment 633, S262{S266. time and spatial distribution of neutron- ux spectral density out- URL https://www.sciencedirect.com/science/article/pii/ side the Russian segment of the International Space Station based S0168900210013732 on data from the BTN-Neutron space experiment. Cosmic Re- Jakubek, J., Cejnarova, A., Holy, T., Pospisil, S., Uher, J., Vykydal, search 55 (2), 110{123. Z., 6 2008. Pixel detectors for imaging with heavy charged URL http://link.springer.com/10.1134/S001095251702006X particles. Nuclear Instruments and Methods in Physics Research Llopart, X., Ballabriga, R., Campbell, M., Tlustos, L., Wong, Section A: Accelerators, Spectrometers, Detectors and Associated W., 10 2007. Timepix, a 65k programmable pixel readout chip Equipment 591 (1), 155{158. for arrival time, energy and/or photon counting measurements. URL https://www.sciencedirect.com/science/article/pii/ Nuclear Instruments and Methods in Physics Research Section S0168900208004245 A: Accelerators, Spectrometers, Detectors and Associated Equip- Jakubek, J., Granja, C., Hartmann, B., Jaekel, O., Martisikova, M., ment 581 (1-2), 485{494. Opalka, L., et al., 10 2011. Imaging with secondary radiation in URL http://linkinghub.elsevier.com/retrieve/pii/ hadron therapy beams with the 3D sensitive voxel detector. In: S0168900207017020 2011 IEEE Nuclear Science Symposium Conference Record. IEEE, Llopart, X., Campbell, M., Dinapoli, R., San Segundo, D., Pernig- pp. 2281{2284. otti, E., 10 2002. Medipix2: A 64-k pixel readout chip with 55-m URL http://ieeexplore.ieee.org/document/6153862/ square elements working in single photon counting mode. IEEE Jakubek, J., Granja, C., Jakel, O., Martisikova, M., Pospicil, S., Transactions on Nuclear Science 49 (5), 2279{2283. 10 2010. Detection and track visualization of primary and sec- URL http://ieeexplore.ieee.org/document/1046904/ ondary radiation in hadron therapy beams with the pixel detector Love, J. J., Pulkkinen, A., Bedrosian, P. A., Jonas, S., Kelbert, A., Timepix. In: IEEE Nuclear Science Symposuim & Medical Imag- Rigler, E. J., et al., sep 2016. Geoelectric hazard maps for the ing Conference. IEEE, pp. 1967{1969. continental United States. Geophysical Research Letters 43 (18), URL http://ieeexplore.ieee.org/document/5874118/ 9415{9424. Kataria, D. O., Fazakerley, A., Green, L., Mathews, S., Hu, H., URL http://doi.wiley.com/10.1002/2016GL070469 Hailey, M., et al., 2013. In-situ sensors and instrumentation for Marti s kov a, M., Jakubek, J., Granja, C., Hartmann, B., Op alka, L., Space Weather and the Sunjammer Solar Sail mission opportunity. Posp  sil, S., et al., 11 2011. Measurement of secondary radiation URL http://www.mssl.ucl.ac.uk/ lmg/spaceweather/Posters/ during ion beam therapy with the pixel detector Timepix. Journal SWAN_Sunjammer_Poster.pdf of Instrumentation 6 (11), C11014{C11014. Kerr, R. A., 5 2013. Planetary exploration. Radiation will make URL http://stacks.iop.org/1748-0221/6/i=11/a=C11014?key= astronauts' trip to Mars even riskier. Science (New York, N.Y.) crossref.524ac17b3fbd199ac514289b41d259a6 340 (6136), 1031. McDonnell, M. D., Tissera, M. D., Vladusich, T., van Schaik, A., URL http://www.ncbi.nlm.nih.gov/pubmed/23723213 Tapson, J., 8 2015. Fast, Simple and Accurate Handwritten Digit Khazanov, G. V., Liemohn, M. W., 5 2002. Transport of photo- Classi cation by Training Shallow Neural Network Classi ers with electrons in the nightside magnetosphere. Journal of Geophysical the Extreme Learning Machine Algorithm. PLOS ONE 10 (8), Research 107 (A5), 1064. e0134254. URL http://doi.wiley.com/10.1029/2001JA000163 URL http://dx.plos.org/10.1371/journal.pone.0134254 Komiya, Y., Shigeyama, T., 9 2017. R-process Element Cosmic Rays Michel, T., Bohnel, M., Durst, J., Sievers, P., Anton, G., 4 2009. from Neutron Star Mergers. The Astrophysical Journal 846 (2), Low Energy Dosimetry With Photon Counting Pixel Detectors 143. Such as Medipix. IEEE Transactions on Nuclear Science 56 (2), URL http://stacks.iop.org/0004-637X/846/i=2/a=143?key= 417{423. crossref.551094144e747e219a0b544ac4779f9b URL http://ieeexplore.ieee.org/document/4812304/ Kotera, K., Olinto, A. V., 1 2011. The Astrophysics of Ultrahigh Mitchell, E. F., Araujo,  H. M., Daly, E., Guerrini, N., Gunes-Lasnet, 20 S., Grin, D., et al., 7 2014. The Highly Miniaturised Radiation Bamford, S. P., Cardamone, C. N., et al., 8 2016. Galaxy Zoo: Monitor. Journal of Instrumentation 9 (07), P07010{P07010. Evidence for rapid, recent quenching within a population of AGN URL http://stacks.iop.org/1748-0221/9/i=07/a=P07010?key= host galaxies. crossref.99e925ebebf12964f35b117711b69246 URL https://arxiv.org/abs/1609.00023 Opalka, L., Granja, C., Hartmann, B., Jakubek, J., Jaekel, O., Mar- Stoe, N., Pinsky, L., Hoang, S., Idarraga, J., Kroupa, M., Jakubek, tisikova, M., et al., 2 2013. Linear energy transfer and track pat- J., et al., 12 2012. Initial results on charge and velocity discrimi- tern recognition of secondary radiation generated in hadron ther- nation for heavy ions using silicon-Timepix detectors. Journal of apy beam in a PMMA target. Journal of Instrumentation 8 (02), Instrumentation 7 (12), C12009{C12009. C02047{C02047. URL http://stacks.iop.org/1748-0221/7/i=12/a=C12009?key= URL http://stacks.iop.org/1748-0221/8/i=02/a=C02047?key= crossref.66ee75f9417ee47019dd26d97f33befe crossref.e678a45dd504d42de05d51a13b68068e Stoe, N., Pinsky, L., Kroupa, M., Hoang, S., Idarraga, J., Am- Parker, B., 2017. Real science, real classrooms. School Science Re- berboy, C., et al., 5 2015. Timepix-based radiation environment view (365). monitor measurements aboard the International Space Station. Parker, B., Fox, E., Rushton, L., 2018. IRIS extending young Nuclear Instruments and Methods in Physics Research Section peoples participation and attainment in STEM and promoting A: Accelerators, Spectrometers, Detectors and Associated Equip- teachers and students as vital members of STEM research ment 782, 143{148. communities. URL https://www.sciencedirect.com/science/article/pii/ URL https://impact.chartered.college/article/ S0168900215001977 parker-iris-stem-students-teachers-participation-research/ Taylor, B., Underwood, C., Dyer, A., Ashton, C., Rason, S., Pav on-Carrasco, F. J., De Santis, A., 4 2016. The South Atlantic Browning, J., 8 2012. The Micro Radiation Environment Mon- Anomaly: The Key for a Possible Geomagnetic Reversal. Fron- itor (MuREM) and SSTL Radiation Monitor (SSTL RM) on tiers in Earth Science 4, 40. TechDemoSat-1. IEEE Transactions on Nuclear Science 59 (4), URL http://journal.frontiersin.org/Article/10.3389/ 1060{1065. feart.2016.00040/abstract URL http://ieeexplore.ieee.org/document/6236273/ Peterson, C., 7 1989. Track nding with neural networks. Nuclear The Pierre Auger Collaboration, T. P. A., 11 2007. Correlation of Instruments and Methods in Physics Research Section A: Ac- the highest energy cosmic rays with nearby extragalactic objects. celerators, Spectrometers, Detectors and Associated Equipment URL http://arxiv.org/abs/0711.2256http://dx.doi.org/10. 279 (3), 537{545. 1126/science.1151124 URL https://www.sciencedirect.com/science/article/pii/ Thomas, S. R., Owens, M. J., Lockwood, M., 1 2014. The 22-Year 0168900289913004 Hale Cycle in Cosmic Ray Flux Evidence for Direct Heliospheric Pierre Auger Collaboration, T. P. A., Aab, A., Abreu, P., Aglietta, Modulation. Solar Physics 289 (1), 407{421. M., Samarai, I. A., Albuquerque, I. F. M., et al., 9 2017. Observa- URL http://link.springer.com/10.1007/s11207-013-0341-5 tion of a large-scale anisotropy in the arrival directions of cosmic Turecek, D., 8 2016. Arrangement of portable display system for rays above 8 1018 eV. Science (New York, N.Y.) 357 (6357), radiation detection. US Patent 9411373B2. 1266{1270. Turecek, D., Holy, T., Jakubek, J., Pospisil, S., Vykydal, Z., 1 URL http://www.ncbi.nlm.nih.gov/pubmed/28935800 2011a. Pixelman: a multi-platform data acquisition and process- Pina, L., Hudec, R., Inneman, A., Cerna, D., Jakubek, J., Sieger, ing software package for Medipix2, Timepix and Medipix3 detec- L., et al., 5 2015. X-ray monitoring for astrophysical applications tors. Journal of Instrumentation 6 (01), C01046{C01046. on Cubesat. Vol. 9510. International Society for Optics and URL http://stacks.iop.org/1748-0221/6/i=01/a=C01046?key= Photonics, p. 951005. crossref.9cbc0e538eeedde66597c91b417135cd URL http://proceedings.spiedigitallibrary.org/ Turecek, D., Pinsky, L., Jakubek, J., Vykydal, Z., Stoe, N., proceeding.aspx?doi=10.1117/12.2185091 Pospisil, S., 12 2011b. Small Dosimeter based on Timepix de- Pinsky, L., Campbell-Ricketts, T., Empl, A., George, S., Tlustos, vice for International Space Station. Journal of Instrumentation L., Turecek, D., et al., 2016. An Update on Medipix in Space 6 (12), C12037{C12037. and Future Plans (Medipix2, 3 &amp; 4). In: 6th International URL http://stacks.iop.org/1748-0221/6/i=12/a=C12037?key= Workshop on Analogue and Mixed-Signal Integrated Circuits for crossref.828425bca11aa9d1c0b4f651ad2115b9 Space Applications. Turner, D. L., Angelopoulos, V., Morley, S. K., Henderson, M. G., Pinsky, L. S., Empl, A., Gutierrez, A., Jakubek, J., Kitamura, H., Reeves, G. D., Li, W., et al., mar 2014. On the cause and extent of Miller, J., et al., may 2011. Penetrating heavy ion charge and outer radiation belt losses during the 30 September 2012 dropout velocity discrimination with a TimePix-based Si detector (for event. Journal of Geophysical Research: Space Physics 119 (3), space radiation applications). Nuclear Instruments and Methods 1530{1540. in Physics Research Section A: Accelerators, Spectrometers, URL http://doi.wiley.com/10.1002/2013JA019446 Detectors and Associated Equipment 633, S190{S193. Underwood, C., Taylor, B., Dyer, A., Hands, A., Ryden, K., Avi- URL https://www.sciencedirect.com/science/article/pii/ son, W., et al., 9 2016. Development of miniaturised radiation S0168900210013549?via{%}3Dihub environment monitors for commercial space ight: Flight results Plackett, R., Tick, T., Ballabriga, R., Vahanen, S., Turecek, D., from the MuREM/RM payloads. In: 2016 16th European Con- Campbell, M., et al., 2010. Current status of the Medipix2, ference on Radiation and Its E ects on Components and Systems Timepix, Medipix3 and Timepix2 pixel readout chips. PoS VER- (RADECS). IEEE, pp. 1{8. TEX2010, 030. URL http://ieeexplore.ieee.org/document/8093160/ URL http://inspirehep.net/record/893444?ln=en United Nations Scienti c Committee on the E ects of Atomic Radi- Potgieter, M., 2013. Solar Modulation of Cosmic Rays. Living Re- ation, 2008. SOURCES AND EFFECTS OF IONIZING RADI- views in Solar Physics 10 (1), 3. ATION UNSCEAR 2008 Report Volume I: SOURCES Report to URL http://link.springer.com/10.12942/lrsp-2013-3 the General Assembly Scienti c Annexes A and B. Tech. rep. Rawat, W., Wang, Z., 9 2017. Deep Convolutional Neural Networks Urban, M., Nentvich, O., Stehlikova, V., Baca, T., Daniel, V., for Image Classi cation: A Comprehensive Review. Neural Hudec, R., 11 2017. VZLUSAT-1: Nanosatellite with miniature Computation 29 (9), 2352{2449. lobster eye X-ray telescope and quali cation of the radiation URL http://www.mitpressjournals.org/doi/abs/10.1162/ shielding composite for space application. Acta Astronautica 140, neco_a_00990 96{104. Rossi, B., Hall, D. B., 1941. Variation of the rate of decay of URL https://www.sciencedirect.com/science/article/pii/ mesotrons with momentum. Physical Review. S0094576516308943?via%3Dihub Smethurst, R. J., Lintott, C. J., Simmons, B. D., Schawinski, K., Vilalta, R., Kuchibhotla, S., Valerio, R., Pinsky, L., 12 2011. 21 Development of Pattern Recognition Software for Tracks of Ionizing Radiation In Medipix2-Based (TimePix) Pixel Detector Devices. Journal of Physics: Conference Series 331 (3), 032052. URL http://stacks.iop.org/1742-6596/331/i=3/a=032052? key=crossref.fce3a0101cb3e64468f2cb40fdeb665c Vykydal, Z., Bouchami, J., Campbell, M., Dolezal, Z., Fiederle, M., Grei enberg, D., et al., 8 2009. The Medipix2-based network for measurement of spectral characteristics and composition of radiation in ATLAS detector. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 607 (1), 35{37. URL https://www.sciencedirect.com/science/article/pii/ S0168900209005956 Whyntie, T., Bithray, H., Cook, J., Coupe, A., Eddy, D., Fickling, R., et al., 10 2015. CERN@school: demonstrating physics with the Timepix detector. Contemporary Physics 56 (4), 451{467. URL http://www.tandfonline.com/doi/full/10.1080/ 00107514.2015.1045193 Whyntie, T., Cook, J., Coupe, A., Fickling, R. L., Parker, B., Shearer, N., 2016. CERN@school: Bringing CERN into the class- room. Nuclear and Particle Physics Proceedings. Whyntie, T., Harrison, M., 3 2015. Full simulation of the LUCID experiment in the Low Earth Orbit radiation environment. Journal of Instrumentation 10 (03), C03043{C03043. URL http://stacks.iop.org/1748-0221/10/i=03/a=C03043? key=crossref.1de52093bd3ca3bc3b66a4b5e0eee571 Whyntie, T., Harrison, M. A., 6 2014. Simulation and analysis of the LUCID experiment in the Low Earth Orbit radiation environment. Journal of Physics: Conference Series 513 (2), URL http://stacks.iop.org/1742-6596/513/i=2/a=022038? key=crossref.d670d88b21563d6f3b39d5b38202c2ac Whyntie, T., Parker, B., 5 2013. Investigating the inverse square law with the Timepix hybrid silicon pixel detector: a CERN@school demonstration experiment. Physics Education 48 (3), 344{349. URL http://stacks.iop.org/0031-9120/48/i=3/a=344?key= crossref.2fdb547cb0a30dd03c059ba251f4e8c0 Willett, K. W., Galloway, M. A., Bamford, S. P., Lintott, C. J., Masters, K. L., Scarlata, C., et al., 2 2017. Galaxy Zoo: morpho- logical classi cations for 120000 galaxies in HST legacy imaging. Monthly Notices of the Royal Astronomical Society 464 (4), 4176{4203. URL https://academic.oup.com/mnras/article-lookup/doi/ 10.1093/mnras/stw2568 Williams, D. J., Mead, G. D., 7 1965. Nightside magnetosphere con- guration as obtained from trapped electrons at 1100 kilometers. Journal of Geophysical Research 70 (13), 3017{3029. URL http://doi.wiley.com/10.1029/JZ070i013p03017 Z abori, B., Hirn, A., Deme, S., Ap athy, I., P azm andi, T., feb 2016. Characterization of cosmic rays and direction dependence in the Polar Region up to 88 km altitude. Journal of Space Weather and Space Climate 6, A12. URL http://www.swsc-journal.org/10.1051/swsc/2016006 Zeitlin, C., Hassler, D. M., Cucinotta, F. A., Ehresmann, B., Wimmer-Schweingruber, R. F., Brinza, D. E., et al., 5 2013. Mea- surements of energetic particle radiation in transit to Mars on the Mars Science Laboratory. Science (New York, N.Y.) 340 (6136), 1080{4. URL http://www.ncbi.nlm.nih.gov/pubmed/23723233 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Astrophysics arXiv (Cornell University)

First results from the LUCID-Timepix spacecraft payload onboard the TechDemoSat-1 satellite in Low Earth Orbit

Astrophysics , Volume 2018 (1810) – Oct 30, 2018

Loading next page...
 
/lp/arxiv-cornell-university/first-results-from-the-lucid-timepix-spacecraft-payload-onboard-the-ZAyoD4cbE3
ISSN
0273-1177
eISSN
ARCH-3330
DOI
10.1016/j.asr.2018.10.045
Publisher site
See Article on Publisher Site

Abstract

The Langton Ultimate Cosmic ray Intensity Detector (LUCID) is a payload onboard the satellite TechDemoSat-1, used to study the radiation environment in Low Earth Orbit (635km). LUCID operated from 2014 to 2017, collecting over 2.1 million frames of radiation data from its ve Timepix detectors on board. LUCID is one of the rst uses of the Timepix detector technology in open space, with the data providing useful insight into the performance of this technology in new environments. It provides high-sensitivity imaging measurements of the mixed radiation eld, with a wide dynamic range in terms of spectral response, particle type and direction. The data has been analysed using computing resources provided by GridPP, with a new machine learning algorithm that uses the Tensor ow framework. This algorithm provides a new approach to processing Medipix data, using a training set of human labelled tracks, providing greater particle classi cation accuracy than other algorithms. For managing the LUCID data, we have developed an online platform called Timepix Analysis Platform at School (TAPAS). This provides a swift and simple way for users to analyse data that they collect using Timepix detectors from both LUCID and other experiments. We also present some possible future uses of the LUCID data and Medipix detectors in space. Keywords: Low Earth Orbit, Space Radiation, Trapped Radiation, Cosmic Rays, Timepix, South Atlantic Anomaly 1. Introduction of sources, from solar (typically E < 10 eV, including Solar Energetic Particles) to within the Milky Way Cosmic radiation consists of high energy particles 9 15 (typically, 10 eV < E < 10 eV Galactic Cosmic Rays), produced by a variety of extra-terrestrial sources. In 15 to extragalactic (typically 10 eV < E), with the source general, cosmic radiation falls into three categories 19 of the highest energy particles (10 eV < E) still heavily depending on their source; Galactic Cosmic Rays (GCRs) debated (see Greisen, 1966, Pierre Auger Collaboration which originate outside the solar system, Solar Energetic et al., 2017, Bell et al., 2018 and many more). See Ferrari Particles (SEPs) which come from the sun, and charged and Szuszkiewicz (2009), Blasi (2013), Deligny and O. particles that are trapped by the Earth's magnetic eld. (2016) and Amato and Blasi (2017) for recent reviews of the study of cosmic rays. When the particles are detected directly they are known as primary particles; particles produced by an Cosmic rays are of great scienti c interest for a variety interaction between a primary particle and some ob- of reasons. In general, (secondary) cosmic rays make up structing medium (e.g. hitting the Earth's atmosphere) around 10% of the background radiation we experience are known as secondary particles. Primary cosmic rays on earth (United Nations Scienti c Committee on the are made up of a variety of particles (protons, electrons, E ects of Atomic Radiation, 2008). High energy rays gammas, light nuclei); an even wider range of particles are have long been a probe of fundamental physics, from the typically produced in secondary particle showers, includ- famous example of the e ect of time dilation on muon ing neutrons, Minimum Ionising Particles (MIPS), which decay rate (Rossi and Hall, 1941, Frisch and Smith, 1963), are usually muons, and pions. Primary cosmic rays span to more recently the new physics suggested by a lack 4 20 a vast range of energies ( 10 10 eV - in contrast the of anisotropy in the cosmic ray electron-positron ratio - maximum energy reached by the Large Hadron Collider potentially from dark matter annihilation (Aguilar et al., is  10 eV, CMS Collaboration 2017) from a large range 2013). High-energy cosmic rays can give valuable infor- mation about high-energy astrophysics, e.g. shockwaves in supernovae (Giuliani et al., 2011, Ackermann et al., Corresponding author 2013), merging neutron stars (Komiya and Shigeyama, Email address: mail@willfurnell.com (Will Furnell) Preprint submitted to Advances in Space Research October 31, 2018 arXiv:1810.12876v1 [astro-ph.IM] 30 Oct 2018 2017) and potentially active galactic nuclei (The Pierre Four of these have been in near continuous operation since Auger Collaboration, 2007, Pierre Auger Collaboration 2012, operated via an onboard laptop. The second use et al., 2017). Cosmic rays can be viewed as complemen- of Medipix in space was on the European Space Agency tary messengers in multi-messenger astronomy, alongside (ESA) Proba V mission, launched on the 7th May 2013 photons, neutrinos and gravitational waves (Branchesi, to Low Earth Orbit (LEO) with an altitude of 820km, 2016). See Ginzburg (1996), Kotera and Olinto (2011) with the spacecraft payload Space Application of Timepix and Castellina and Donato (2013), for recent reviews Radiation Monitor (SATRAM, Granja et al., 2016) of the role of cosmic rays in astrophysics. More locally onboard. SATRAM carries a single Timepix detector and cosmic rays are a probe of solar physics (e.g. see Potgieter, is operating and continuously taking data today. In addi- 2013), and are a key component of `space weather' (e.g. tion, Exploration Flight Test 1 (EFT-1), the rst ight see Turner et al., 2014). Space weather can have an of the Orion Multi-Purpose Crew Vehicle (MPCV) on a impact both on ground based communications systems two orbit, 4.5 hour trip on the 5th December 2014 took (e.g. a modern day Carrington event, see Love et al., Medipix data the farthest from Earth to date at 5910 2016) satellites and spacecraft electronics (Choi et al., km (Gaza et al., 2017). Most recently, on the 23rd June 2011) and organisms (including humans). Understanding 2017 the cubesat VZLUSAT-1 (altitude 510km, D aniel the radiation environment is vital for understanding the et al., 2016, Urban et al., 2017) carried a miniaturised impact of the dose astronauts receive on the International x-ray telescope, that uses Timepix detectors (Baca et al., Space Station (ISS, e.g. the link between received 2016), into orbit for astrophysical, space weather studies, dose and susceptibility to cataracts, Cucinotta et al., and terrestrial X-ray monitoring applications (see Pina 2001, see Cucinotta, 2007 for an overview), or on a hy- et al., 2015). pothetical voyage to Mars (Kerr, 2013, Zeitlin et al., 2013). In this paper we report the rst results from one of The development of novel energetic particle detector the early uses of Medipix in orbit (and the rst on a technologies at the European Organization for Nuclear commercial platform, and the rst with Medipix detectors Research (CERN) provides an opportunity to improve in a 3D con guration), the Langton Ultimate Cosmic ray measurements of cosmic rays both in space and on the Intensity Detector (LUCID) on board TechDemoSat-1. ground. The turn of the 21st Century saw the advent of The dominant source of particles detected by LUCID are photon counting pixel detectors for radiation detection trapped electrons and protons, and the instrument is de- with the development of Medipix detectors (Bisogni et al., signed such that it could be deployed as a hosted payload 1998, Campbell et al., 1998). Medipix detectors (now for satellite environmental monitoring. We discuss the de- entering their fourth generation, Medipix1 Amendolia sign of the instrument and its operations, a new platform et al., 1999, Medipix2 Llopart et al., 2002, Medipix3 TAPAS for managing the large amounts of data produced Ballabriga et al., 2011, Medipix4 collaboration founded by the experiment and a new machine learning algorithm 2016) can detect and di erentiate between many types for the automated classi cation of particles which can be of ionizing radiation, and present many advantages used for other Medipix applications. The rst results and compared to other methods - but at the cost of a small some early science applications (e.g. mapping out the collecting area and comparatively high expense. Medipix South Atlantic Anomaly) are also presented. In addition, detectors have been used in a wide range of applications, LUCID is linked to a extensive programme of education including high-energy physics experiments (Grei enberg and research in the classroom, CERN@school, where et al., 2009, Vykydal et al., 2009, Collins et al., 2011), students can use Timepix detectors for both novel tests medical physics (hence `Medi', e.g. Blanchot et al., 2006, of traditional classroom experiments (e.g. inverse square Butzer et al., 2008, Marti s kov a et al., 2011, Jakubek law, see Whyntie and Parker, 2013) as well as original et al., 2011, Hartmann, 2013) and small animal imaging science, for example the Radiation In Soil Experiment (Accorsi et al., 2008). (RISE) which has measured the radiation in di erent geological samples across the UK, and construction of More recently, there has been increased interest in a robotic three-dimensional radiation scanner (Whyntie their application in space (Kroupa et al., 2015, Granja et al., 2016, see also Colthurst et al., 2015, Parker, 2017, et al., 2016, Gaza et al., 2017, Urban et al., 2017). In Parker et al., 2018 for further general information). All of particular their ability to distinguish between di erent these projects are managed through the same data storage particle types and give angular information have, for and reduction pipeline used in this work, see section 3.2. example, proved valuable in understanding the radiation Cubesats have already been shown to be highly e ective environment of the ISS. Using Medipix in space was rst educational tools e.g. Li et al. (2013), and other Medipix discussed in Pinsky et al. (2011). Seven NASA/IEAP- devices have been used in the CERN@school programme developed Radiation Environment Monitors (REMs), e.g. RasPIX Turecek (2016). Timepix detectors in compact USB mounting, have been deployed to the ISS (altitude 400km), see Turecek et al. In addition to LUCID, the ISS REMS, SATRAM (2011b), Kroupa et al. (2015) and Stoe et al. (2015). and the VZLUSAT-1 x-ray telescope, there are also 2 future planned deployments, see Pinsky et al. (2016). et al., 2002, Llopart et al., 2007; Ballabriga et al., 2011). Future missions include; a particle telescope architecture The detectors used are equipped with a 300m silicon containing two Timepix detectors in sync, on the Rapid sensor. The Timepix chips contain 256 256 pixels, each International Scienti c Experiment Satellite (RISESAT, measuring 55m on each side, giving a total collecting Kuwahara et al., 2011, Granja et al., 2014a), a Japanese area of 1.98cm . The circuitry to digitise the output of FIRST mission to orbit at 700km, further Medipix being each pixel is contained within the footprint of the pixel, sent to the ISS, HERA monitors (units containing single meaning only digital information is transferred out of Timepix detectors being developed at NASA for use on the pixels. Timepix can operate in event counting mode future MPCV missions), the proposed Biosentinel astro- (the base functionality of Medipix2, where a counter is biology deep-space cubesat mission and on trans-lunar incremented each time that during the shutter time period NASA-ORION missions in the 2020s. the charge deposited is over the designated threshold), The paper is organised as follows. In section 2 we arrival time mode (where essentially the rst time during describe the LUCID payload. In Section 3 we describe the shutter time period that the charge over a threshold the data reduction process, using GridPP resources, is deposited is recorded) and energy sensitive Time over processing all the data that was produced by the payload Threshold (ToT) mode (measuring the signal amplitude with a new machine learning analysis framework. We providing the per-pixel deposited energy). LUCID was also describe the development of the Timepix Analysis always run in ToT mode. Platform at School (TAPAS), a web platform to present the reduced LUCID data to secondary school students. In Section 4 we give the preliminary results that we have 2.2. LUCID Instrument Design gathered based on this data. In Section 5 we discuss our We brie y summarise the instrument design here; a results and future planned work. Finally, in Section 6, we full technical overview and detailed design information of summarise our ndings. LUCID is in the LUCID System Design Document (D. Cooke, SSTL, private communication). The payload has ve Timepix radiation detectors in a cube-like con gu- ration (see Fig. 1b), with four detectors orthogonally 2. The LUCID Payload positioned facing outwards (TPX0 through TPX3), and 2.1. LUCID and TechDemoSat-1 the fth in the centre (TPX4), facing outwards (relative to the centre of LUCID). A photograph of the instrument LUCID is a payload on the technology demonstra- is shown in Figure 1. The chips were surrounded by tion satellite TechDemoSat-1 (TDS-1, see gure 1). a 0.75mm thick aluminium dome which blocks intense The project started in 2008, and was developed as a light, plasma and low-energy charged particles. LUCID is collaboration between Langton Star Centre secondary mounted on the `Earthside' of TDS-1. school student researchers, the Medipix Collaboration, and Surrey Satellite Technology Limited (SSTL), who The detectors were calibrated by the Institute of built both LUCID and TDS-1. LUCID is part of the Experimental and Applied Physics (IEAP) at the Czech TDS-1 Space Environment Suite, which consists of the Technical University in Prague. The calibration process Miniature Radiation Environment and e ects Monitor involves exposing the detectors to 'X rays of discrete (MuREM, Taylor et al., 2012, Underwood et al., 2016), energy, and modelling the low energy end non-linear the Charged Particle Spectrometer (ChaPS, Kataria response of each individual pixel, see Jakubek et al. et al., 2013) and the Highly Miniaturized Radiation (2008) and Jakubek (2011). Monitor (HMRM, Mitchell et al., 2014, Guerrini et al., 2013). TDS-1 launched on 8 July 2014 (15:58:28 UTC) on The performance and expected measurements of a Soyuz-2-1b launch vehicle with Fregat-M upper stage LUCID were simulated in Whyntie and Harrison (2014), from the Baikonur Cosmodrome in Kazakhstan, into a Whyntie and Harrison (2015), who simulated the en- 635 km, 98.4 Sun-synchronous orbit. LUCID began vironment of LUCID and expected data rates using data collection shortly after launch, and data collection European Space Agency's (ESA) SPace ENVironment ceased on the 4th July 2017. TDS-1 operations have now Information System (SPENVIS, Heynderickx et al., 2004) ended, and at some point in the medium-term it will be and GEANT4 (Agostinelli et al., 2003, Allison et al., 2006 deorbited by the Icarus-1 Cran eld Drag Augmentation and Allison et al., 2016). System de-orbiter (Hobbs et al., 2013) which will over the next 25 years guide the spacecraft into the Earth's 2.3. LUCID Operations atmosphere, where it will disintegrate. Although the payload has ve detectors, for the The detectors used in LUCID are based on the Timepix majority of its time in orbit, only three were taking data ASIC chip (Llopart et al., 2007, Plackett et al., 2010), part at any one time. This was because LUCID would be of the second generation of Medipix (Medipix2, Llopart drawing too much power with all ve detectors taking 3 Figure 1: LUCID shown integrated on TDS-1 in the SSTL assembly room b) the LUCID payload showing the Timepix detector array (taken from the LUCID System Design Document). The four orthogonally positioned detectors are visible on the right hand side of the instrument. Table 1: Payload Speci cations Detectors 5 Timepix hybrid pixel detectors (300m silicon), 1 RADFET (metal-oxide-silicon) Power 1 permanent 5V supply and 1 28V supply for the Timepix detectors. With all 5 detectors running, maximum power usage is 8W Storage 2GB NAND ash Data transfer Maximum 20Mbit/s via X-Band Physical Dimensions 220mm x 135mm x 33mm (SSTL nano tray) Mass 1.2kg Dome 0.75mm aluminium, energies excluded; E > 0:4MeV and E > 10:0MeV (Whyntie and e p Harrison, 2015) 4 The nal data product from LUCID consists of 256x256 Table 2: Timepix con guration DAC-parameters, used for a large PNG images of each frame, CSV les that include the par- number of runs on all detectors (ID 321) ticle counts on a per frame basis, x,y,C formatted les for Active detectors 0, 1, 2, 3, 4 each frame, together with associated metadata and navi- Frame rate 1 per second gation and time stamp. This is post-conversion, as data Shutter Exposure Time 0.3 s was retrieved in a proprietary compressed raw format. Clock 33 MHz Number of frames 100 Bias Voltage 20.02 V 3. LUCID Data Analysis IKrum 1 Disc 127 Di erent particles typically produce very di erent Preamp 255 tracks when they pass through a Medipix detector (e.g. DAC Code 6 Llopart et al., 2002 and Bouchami et al., 2011). Given Sense DAC 0 the large volume of data collected by the instrument, it is Ext. DAC Sel. 0 necessary to automate the detection of di erent particles. Bu AnalogA 127 In this section we present two new high-performing Bu AnalogB 127 machine learning algorithms for classifying tracks in Hist 0 LUCID frames into particle types. This is a dicult THL Fine 300 task both because of the large quantity of data (meaning THL Coarse 7 that the algorithm has to be reasonably fast, and the VCAS 130 task cannot be done solely by human classi ers), as FBK 128 well as the general diculty in classifying tracks (essen- GND 80 tially a pattern recognition problem e.g. Holy et al., 2008). CTPR 0 THS 100 Firstly a clustering algorithm is run on the LUCID BiasLVDS 128 frames to identify individual particle tracks. After having RefLVDS 128 identi ed the track for each individual particle, we consid- ered two main methods classifying particles: data. Towards the end of operations all ve detectors were `Metric-Based Network' (MBN): we extract a given switched on and taking readings as the other payloads set of features for each particle track and then clas- entered di erent phases of operation. The detectors are sify the particle based on these metrics using ma- built and con gured to operate in sync - i.e. they capture chine learning. This network does not take per-pixel frames simultaneously. Speci cations are shown in table 1. energy distribution into account. `Deep learning': to use convolutional neural net- Each detector is able to independently capture a works to classify the particles tracks directly with `frame' of the radiation that passes through the detector a pure deep learning classi cation algorithm. This over some time period. LUCID was run with a range of network does take per-pixel energy distribution into frame rates and shutter exposure times, with the most account. commonly used shown in table 2, and a frame taken every second. The payload has a 2GB NAND ash storage Both of these methods are supervised machine learn- which is used to store data between passes of the ground ing algorithms and therefore the main limiting factor for stations, allowing LUCID to take data for more extended the development of the classi cation algorithms is the periods of time. The amount of available storage space computing power available and the amount of labelled and data transfer requirements of other experiments data that had been collected. In this work the training dictated frame rate. data is labelled by human classi ers. LUCID operated in a data gathering capacity from We also compare the performance of our algorithms to late October 2014 until July 2017. However, the rst an analytic classi er (i.e. tracks are classi ed based on an few months of operation were dedicated to instrument analytic function of the feature metrics) used in an early commissioning, and were used to nd the optimal con g- stage of the analysis of the data (see Whyntie et al., 2015). uration settings for data recording and the abilities of the payload. Nominal operations commenced in April 2015. Our approach can be compared to other studies of During this time, over 2.1 million frames were captured, pattern recognition and cluster analysis in Timepix de- over 82 runs and 11,700 les. A run is an 8 day capture tectors, such as Vilalta et al. (2011), Hoang et al. (2012), period, de ned by the operational schedule of TDS-1. Opalka et al. (2013) and Holy et al. (2008). Other possible machine learning approaches to classifying particles not 5 considered here include generative adversarial networks This allowed the latitude and longitude of LUCID when as well as autoencoders. the capture took place to be calculated. The classi cation was then run on all frames in the particular le, and for each frame alpha, beta, X-ray, proton, muon and other 3.1. Technical Implementation particle classi cations, and the latitude, longitude, frame number, capture timestamp were submitted directly to the The payload was operated by submitting Payload TAPAS (section 3.2) database via a POST request to a Task Request (PTR) les to SSTL. These les con- REST API endpoint. Each run and le (with its times- tain information about the start time of the run, the tamp, ID and con guration le used) was also submitted con guration le to use, and the overall schedule for this way. capturing data in this run. The LUCID con guration les specify which detectors should be used in this 3.2. The TAPAS Data Analysis and Visualisation Tool capture and the settings of each detector to be used, for instance the shutter exposure time and threshold value. We have developed The Timepix Analysis Platform at School (TAPAS) to allow secondary school student During the operating lifetime of the payload, data researchers across the UK to analyse and share the data was transferred from the satellite to SSTL and then that they gather using Timepix radiation detectors across downloaded to an IRIS server at the Langton Star all CERN@school projects, and additionally as a home Centre in Canterbury, UK. A cron job (a system used for for the particle count data from the LUCID experiment. scheduling jobs on UNIX based systems) ran every day The platform allows users to upload their own data, taken to check for the existence of new les on the SSTL FTP with the Timepix detectors using a software package server, and if les were detected, they were downloaded called Pixelman (Turecek et al., 2011a), or data which and processed immediately, converting them from a has been provided to them, such as the TimPix ISS-REM proprietary format for LUCID into individual frames of radiation data. A web application was chosen instead of x, y, C les (x-coordinate, y-coordinate, ToT value; tab a desktop application because it is very easy to access - separated). Metadata (time of capture, detector frame students only need a web browser and internet connection was from, the le the frame belongs to and location to use the service. information) was stored in an SQLite database. Once a user has uploaded a dataset, a cron job is used Pre-processed data was compressed and uploaded to run the analysis service every 10 minutes to process to GridPP (GridPP et al., 2006, Britton et al., 2009) the data. The analysis service uses multiple processes to storage, using the CernVM (Buncic et al., 2010). At the analyse large amounts of the data in a parallel fashion, same time as the uploading process, custom software was using the lucid utils LUCID algorithm. The service also developed to process runs in parallel, allowing many runs generates an image of every frame processed. TAPAS has to be analysed at the same time, exploiting the highly also been used to analyse small amounts of the LUCID parallel nature of running jobs on GridPP. data (the whole LUCID data set required the resources of GridPP). The processing software has been developed in Python, using Tensor ow (Abadi et al., 2016) machine learning for TAPAS is a Django framework based web application, particle analysis (see sections 3.5 and 3.6). The CernVM primarily written in Python for the page generation and File System (CVMFS) was used to deploy the software HTML, CSS and JavaScript for the frontend user facing dependencies and Python interpreter to the grid worker elements. The web platform has a MariaDB database nodes to run the software, and jobs were submitted backend that is used to store all metadata relating to a using the gLite middleware to a speci c GPU-backed job LUCID run, le and frame and particle classi cations. queue. The full source for the software is available at: This database also includes information and analysis https://github.com/willfurnell/lucid-grid/. results relating to user-made uploads. Each job running on a worker node copied data from The platform includes an API, using the `Django Rest a storage element to its working directory, extracted it Framework' to allow uploads and particle classi cations and then ran the classi cation algorithm. This process to be submitted using a programmatic REST (Represen- included getting frames and their metadata (namely the tational state transfer) interface. This is of particular capture time) from the pre-process database, and then cal- importance to the LUCID data, as this is how particle culating which of the Two Line Element les (TLE, les classi cations were submitted to the database from the that contain the information to track the location of the satellite ) had a date closest to the frame's timestamp. lines of data which can be used together with NORAD's SGP4/SDP4 orbital model to determine the position and velocity of the associated satellite.' - https://celestrak.com/columns/v04n03/ `A NORAD two-line element set consists of two 69-character 6 (typically IRIS secondary school researchers) to simply click through automatically generated questionnaires. The responses would be stored in a database alongside the pixel data and the metadata for nding out which frame the cluster belonged to. The LUCID Trainer web application is written in PHP for dynamic page generation and HTML, CSS and JavaScript for the frontend user interface. All the LUCID data is accessed via a custom REST API connecting the LUCID metadata database to the web application. The user response data is sent via an AJAX request to another PHP le that then stores the response in a simple SQLite database which was used due to its lightweight requirements. The training set used consisted of 1800 particles Figure 2: An example LUCID frame with individual particle tracks (tracks identi ed as in section 3.3). These tracks were identi ed. each classi ed once by a student researcher, with 24 classi ed as alpha particles, 988 as beta/electrons, 547 software running on GridPP worker nodes. as X-ray/photons, 27 as muons, 160 as protons and 54 as `other'. Clusters that the user could not identify, or We have programmed TAPAS to allow students to overlapping clusters were those classi ed as `other'. The download CSV les with particle count data, allow- student researcher classi ers themselves were trained ing them to conduct further analysis using their own using example tracks with known particle type, and choice of software packages - or even by choosing to reference diagrams similar to Figure 1 in Bouchami et al. write their own, in a programming language they are (2011) and Figure 1 in Whyntie et al. (2015). The student most comfortable with using. All of the LUCID par- researchers, and therefore the algorithm, may have had ticle count data is downloadable as CSV les from TAPAS. issues separating one particle type from another. The resulting training set is thus based on human classi ca- tions, which as a methodology is necessarily less accurate 3.3. Event Identi cation than using a training set constructed from known cali- brated sources. Nevertheless, these preliminary initial The rst step of classifying particles in a given Timepix results still give us a good overview of the distribution detector frame is to identify each individual particle track. of morphology of detections, and in the future calibrated The frame is fed in as a 256 x 256 matrix of energy values. classi cations can be generated and used for training to Non-zero pixels are systematically selected, and any obtain more realistic particle counts. non-zero pixels in the `ring' of the 8 adjacent (including diagonals) pixels are then added to the associated cluster. The nature of supervised machine learning algorithms For each adjacent pixel, all other adjacent pixels are trained on human classi cations means that the algorithm checked until there are no more pixels with non-zero can at best reproduce the classi cation that a human energy values. The clustering algorithm works through classi er would give. Similar projects such as Galaxy Zoo the entire frame and returns a list of clusters, each of (e.g. Lintott et al., 2008, Willett et al., 2017, Smethurst which is a list of pixels. An alternative clustering method et al., 2016) and the rest of the Zooniverse have shown is discussed in Granja et al. (2018), which uses track success in the classi cation of large samples of image shape and deposited energy values to classify the particle and image-like data based on a training set of human type, energy, and direction of events. Figure 2 shows an example LUCID frame where each particle track has been classi cations. Vilalta et al. (2011), in contrast to this successfully identi ed using the clustering algorithm. work, trained their algorithm using data taken at the Heavy Ion Medical ACcelerator Facility (HIMAC) in Chiba, Japan e.g. they red beams of known particle 3.4. Training Data type and energy at Medipix detectors, so they had a set of tracks labeled by the true particle properties (see also Both algorithms considered here are supervised, and Jakubek et al., 2010). Our human labelling of tracks will require that a subset of the data is labelled with the `true' be imperfect - however human labelling is still a valuable classi cation. To generate training data for classifying approach to develop for the classi cation of tracks as particle tracks, a web application called LUCID Trainer human classi ers can identify `unexpected' tracks that was created (Figure 3). It allows volunteer classi ers might be in the LUCID data but not in a laboratory 7 Figure 3: Screenshot from volunteer classi cation questionnaire, as seen by the student researchers. produced training set (e.g. Beck et al., 2018 found that a  Density - This is calculated using the calculate den- combination of human classi cation and machine learning sity function. classi cation gave the best results). Line Residual - This is calculated using the calculate non-linearity function. 3.5. Metric Based Network Curvature Radius and Circle Residual - These are The Metric Based Network (MBN) approach to calculated using the nd best t circle function. classifying the particles was to calculate a small number Average Neighbours - This is calculated using the of easily computed features for each track, and to then nd average neighbours function. classify the particles based on the metrics that had been calculated. Width - Number of Pixels / Diameter (if number of pixels > 1 else width = 0) The primary classi er used was the multi-layer neural network. The hyperparameters (number of hidden layers, The precise de nitions of these features are given in number of nodes in each hidden layer etc.) for this Appendix B. In early analyses of the data these metrics architecture was optimised manually. Other machine were used for an analytic classi cation of the tracks (e.g. learning classi ers were used for comparison as well as for if number of pixels less than 4, then classify as a X-ray ensuring that the limiting factor for the accuracy was the etc.), also described in Appendix B. size of the training set and not the architecture. The nal multi-layer neural network architecture has The general principles behind the multi-layer neural 8 inputs in the input layer (for the 8 metrics), 128 nodes network are described in Haykin (1998), and see the in the rst hidden layer with the sigmoid activation extensive literature on similar problem of classifying function, 48 nodes in the second hidden layer with the handwritten digits (a similar problem) e.g. McDonnell sigmoid activation function, and 6 outputs in the output et al. (2015). See Denby (1988) and Peterson (1989) layer with the softmax activation function, see gure 4. for early uses of neural networks in studying particle No dropout or regularisation was needed as the size of tracks in high energy density physics experiments, and the training dataset was already limited. The Gradient Farrell et al. (2017) for a more recent example in the LHC. Descent Optimiser was used as the training algorithm. The network was trained with a batch size (the number The eight metrics calculated from the pixel cluster rep- of training examples in one forward/backward pass) of resenting the particle track for the algorithm are: 128 particles. Additional classes can be added to the Metric-Based Network with ease as the outputs are Number of Pixels - This is calculated as the length one-hot encoded however it will have to be trained from of the pixels list. scratch on the dataset. Radius - This is calculated using the calculate radius function. 8 Figure 4: The architecture of the SML neural network used (diagram not to scale) 3.6. Deep Learning Deep learning algorithms present the possibility to Figure 5: Overall accuracy of each algorithm classify the particle tracks directly from the images (which contain more information) rather than from the tree and random forest, all available in Tensor ow) were pre-identi ed features considered in the MBN algorithm. also tested to con rm that our network architecture was Deep learning approaches have high requirements of optimal but are not shown in the gure. The hyperpa- computing power and typically are used with GPUs, so rameters for these algorithms were not investigated and we only present some early results here, although the the default parameters were used. approach could lead to much higher levels of accuracy with future work. Figure 5 shows the percentage correctly classi ed for a) the Metric-Based Network (MBN), b) a Deep Convolu- The deep learning architecture considered was a Deep tional Network (DCN) and c) the Analytic Classi er (AC). Convolutional Neural Network (CNN e.g. see Rawat and Wang, 2017 for a recent review) run for 5000 epochs. Epochs are one forward pass and one backward pass of all All performed between 80-95% accuracy, with the the training examples. We do not describe the algorithm MBN performing best on the test data. The AC had in great depth as the method remains preliminary. A a high accuracy of 84% but still less accurate than the Deep Residual Network (DRN e.g. He et al., 2016) has MBN. This is because the dataset consists of much also been tested on the data, but as of yet only for a small more beta and X-ray particles and therefore the overall number of epochs so has not yet attained a comparable accuracy appears to be high while its performance on a particle-by-particle basis was much poorer (see Fig high level of precision, but may be competitive in the 6b). Figure 6 shows the confusion matrices for our best future. algorithm, the Metric-Based Network, and the Analytic Classi er to show how machine learning leads to much Other advantages of the Deep Convolutional Network better performance. Each square shows the probability to the Metric-Based Network is that the Deep Convolu- one type of particle has of being classi ed as another. tional Network allows the energy values to be considered Complete success in classi cation would correspond to a (the feature metrics use for the MBN algorithm don't diagonal. The Metric-Based Network algorithm performs use the energy value of the pixels, just whether the pixel well, with the only substantial misclassi cations being was non-zero or not) and the convolutional lters allow 30-40% of muons and protons being misclassi ed as humans to visualise what the network has learned. electrons - partially because of similarities in the track shapes, partially because electrons dominate the overall sample, and partially because the labels used in the 3.7. Algorithm Performance training are imperfect. The Analytic Algorithm performs Figures 5 and 6 show the performance of the al- much poorer, misclassifying all particles as either electrons gorithms, the Metric Based Network (MBN) Deep and protons a substantial fraction of the time, and failing Convolutional Network (DCN) and the Analytic Classi er to identify any muons or `Others'. (AC). Other `o the shelf ' machine learning classi ers (support vector machines, k-nearest neighbours, a decision 9 Figure 6: Confusion matrices for the analytic classi er and MBN. Squares in the grid are colour coded by percentage of `actual class' classi ed as `predicted class'. 60 W) and the poles (roughly 60 N, 60 S) are clearly Table 3: Event Classi cations (all frames) evident for all particle types. We also show in gure 8 the Alpha 75902 ratio of heavy charged to light charged particles - this is Beta 169391026 non-uniform, showing that that heavy charged particles X-ray 369816217 have disproportionate intensity in the SAA and poles Proton 1065545 compared with the light charged particles. Muon 569771 Other 6703109 The aim in this section is to illustrate that LUCID is giving sensible results, as opposed to give detailed measurements of dose, linear energy transfer (LET) 4. Results and particle energy spectra. Although LUCID can In this section we present some preliminary results from make estimates of the energies of particles because the the data and particle classi cations described in Section 3. detectors have been calibrated, this analysis is beyond We use the classi cations from the Metric-Based Network the scope of this study and will be the focus of future work. algorithm e.g. the particles are classi ed only based on track morphology, and some tracks will be in incorrectly To create ux maps, a speci c date range was cho- classi ed, both due to the algorithm not having 100% ac- sen that covered the whole Earth, while also having a curacy (see Fig 6b) and imperfect labelling in the training homogeneous set of con guration les (also limiting the set. e ects of any time evolution of the radiation eld). The sub-set of the data considered consisted of 404 les, from 4.1. Classi cations 2016-08-17 to 2016-09-21, where we took the rst 10 frames from each of the les to result in a total of 4040 Table 3 presents the particle classi cations that we frames. For each le, the measured ux (count rate per have obtained for the whole lifetime of the LUCID pay- unit time per unit collecting area) was calculated. Then load. Di erent con guration les may produce di erent a K-dimensional (K-D) tree (with a maximum number classi cations due to saturating the frame (or conversely of neighbour lookups of 100) was used, so that for every having completely empty frames), and therefore these re- given latitude and longitude reached by LUCID, the sults should be taken as a preliminary analysis, rather than plotted ux at that point is the average of neighbouring con rmed results. Full payload data results are available frames ux, weighted inversely by the distance to the on TAPAS https://tapas.researchinschools.org frames. 4.2. Radiation Map and the South Atlantic Anomaly We plot in gure 7 the number of particles detected 4.3. Heavy Charged Particles over the Earth's surface for 4000 frames for di erent Of our tracks, about 3% were classi ed as `Other', and particles. Higher radiation levels (by more than a factor a substantial proportion of these particles are likely nuclei of ten) around South Atlantic Anomaly (an area of heavier than helium. We present example tracks in gures known increased radiation ux, centred at roughly 30 S, 10 Figure 7: Radiation maps for alpha particles, electrons, X-Rays, protons, muons and unclassi ed particles 11 Figure 8: A map of the ratio (Betas + Muons+1) / (Protons + Alphas+1). The +1 on the denominator to prevent division by 0 and +1 on numerator for a 1:1 ratio when both are 0. 12 9a and 9b, showing the extended shape. Timepix are par- 5.1. Other Experiments ticularly well suited for identifying heavy charged particles Our electron/proton percentages appear to be similar as they can extract track shape (as opposed to just particle to other experiments. Our radiation maps are similar to energy), see Granja et al. (2011), Stoe et al. (2012) and those found by the ISS-REMs (e.g. Fig 7. Stoe et al., Hoang et al. (2012). Identifying ux levels of heavy ions 2015). We do however go to higher latitudes (the ISS data carries important astrophysical information e.g. Aguilar only goes up to 55 N/S) mapping out the polar regions. et al. (2018). SATRAM, Granja et al. (2016) also map to the higher alti- tudes. Litvak et al. (2017) nd similar plots in the neutron 4.4. Preliminary Tests ux (measurements from the BTN-Neutron space experi- We do some early geometric tests to con rm that ment on the ISS), one of the particles LUCID can't detect. the experiment is gives reasonable results, as a prelude to future work investigating particle isotropy and the The ability of Timepix to classify heavy charged par- physics of particle transport from the sun and the trapped ticles is highlighted in Kroupa et al. (2015). Our results electron model. add to their ndings that Timepix have potential for use as a `heavy ion' monitor, or even telescope (Branchesi, Rather than travelling in straight lines, the particles 2016). Furthermore, our use of human classi ers provides are de ected towards the poles and the South Atlantic a testbed for future applications where humans are used to Anomaly (SAA) by the earth's magnetic eld, which is classify tracks that a traditional computer algorithm may why the particle count is higher as LUCID gets towards the struggle with. poles. We found slight evidence for marginally higher mea- sured number counts when LUCID is in front of the earth 5.2. Future Plans than behind the earth (relative to the Sun), mean of 7.1 Future work will focus on developing the algorithms particles per frame `dayside', and 6.3 particles `nightside'. and analyses presented in this work. In this paper we Dayside and nightside were de ned by calculating the an- focussed on the classi cation of track morphologies. A gle made by the LUCID satellite, the centre of the earth subsequent paper will explore the calibration of energy and the sun using the latitude, longitude and timestamp of measurements, and make dose maps, measurements of each frame. When this angle was between 0 and 90 , the LET, and particle energy spectra. We will also redo our frame was classi ed as on the `dayside' (i.e. the half of the analysis with larger training sets and improved imple- earth facing the sun) and when it was more than 90 it was mentations of the deep learning approaches. This new classi ed as on the `nightside'. Future work will focus on analysis will use particle tracks that have been validated investigating if this is a systematic in the data or not, and and calibrated from a de ned radiation eld so we can testing our results within current models of the Earth's identify events more accurately. magnetosphere, as the Earth's magnetosphere is known to be di erent on the `day' and `night' sides of the Earth. We will investigate di erences in ux measured in Pressure from the solar wind coming from the sun on the the di erent detectors (e.g. di erences between Timepix dayside compresses the magnetic eld, and on the night- perpendicular to each other). In addition, there is still side elongates it. In the Van Allen radiation belts (which much to exploit from the ability of Medipix detectors LUCID is close to/passes through in the SAA), charged to measure both the elevation and azimuthal angle of particles are trapped within the magnetic eld, and where particle trajectories in space e.g George (2014), Kroupa this eld is condensed on the dayside there will be a higher et al. (2015) nd that particles are highly directional particle density, which is consistent with a higher parti- in the SAA (c.f. Granja et al., 2014b) See also rocket cle count on this side of the earth (e.g. Williams and experiments at lower altitudes e.g. Z abori et al. (2016). Mead, 1965, Domrachev and Chugunin, 2002, Khazanov This would require identi cation of particle track angle and Liemohn, 2002). These results show that the LUCID and inclination as part of the machine learning classi- data can contribute to testing and update models of space cation algorithm, although we note that this must be weather. done carefully to avoid uncertainty regarding the angles of entry. In addition it would also be worthwhile to compare the student researcher track labels to labelled 5. Discussion tracks from known sources - as already mentioned, the In this section we discuss how our results compare to classi cations of Vilalta et al. (2011) and Hoang (2013) comparable space-based space radiation detectors, what have the advantage of `true' labels compared with our our results mean for future use of Medipix in space, and human-classi ed labels, but may nd it more dicult to what our future plans with LUCID data are. identify unexpected tracks. With classi cations of particle inclination as well as particle type we will use the particle angle measurements 13 Figure 9: Sample tracks of likely heavy charged particles. The colour scale is linear and scales with the energy deposited by the ion in the silicon layer for each pixel. to test the isotropy of particle ux in the trapped electron ent types of radiation and the inverse square law). IRIS is model. We will also investigate evolution in ux over currently undertaking pedagogical research about the ef- time, to link measurements to solar cycles (Thomas et al., fectiveness of this approach to teaching, and is developing 2014), and investigate the growth of the SAA and any plans to expand CERN@school to other CERN member possible link to geomagnetic reversal (Pav on-Carrasco states. and De Santis, 2016). 6. Conclusions It has recently been shown that the kinetic energy of particles can be reconstructed using a Timepix detector In this paper we introduce the LUCID detector (the (Kroupa et al., 2018), which will also be a topic of future third use of Medipix detectors in space, the second in open investigation with the information from the data that we space, and the rst on a commercial platform and with have obtained from LUCID. a 3D con guration) that ew aboard TDS-1 taking data from 2014 to 2017. We describe the data pipeline from 5.3. Lessons From CERN@school data collection to reduced catalogue of classi ed particles, We view LUCID as technological test of Medipix a novel machine learning particle track classi cation detectors for future space missions, giving Timepix data algorithm, and some early science results. We also discuss at a di erent altitude to other experiments, for the rst LUCIDs links to the larger CERN@school ecosystem. time on a commercial platform, in a novel con guration of chips. However we also view CERN@school as an The payload was operated by submitting Payload Task important test of the use of Medipix technology. Medipix Requests to SSTL. Data immediately after collection was had previously been predominantly used by professional stored on a ash partition on the spacecraft until passes scientists. CERN@school was the rst widespread use of over ground stations. The data was passed from the Medipix across 100s of institutions, with users ranging ground station to SSTL, and then to IRIS servers. From from novices to experts. The challenges of handling data there the LUCID frame data were passed to GridPP for from a very heterogeneous set of sources for use by a processing, and the reduced data passed back to the IRIS wide variety of users of di erent levels of expertise is servers. a test bed for any future hypothetical large-scale use of Medipix outside of academia and research labs e.g. We use machine learning techniques to classify Medipix as a personal radiation monitor in a nuclear power Timepix particle tracks, presenting two types of al- plant or for nuclear medicine workers (Michel et al., 2009). gorithm: a) a metric-based neural network classi er that uses eight pre-identi ed `features' and b) a deep Alongside the research applications, CERN@school de- learning approach, showing both perform well. These tectors in schools also have an educational role (e.g. as a algorithms perform competitively compared with exist- replacement for a geiger counter in teaching about di er- ing analytic algorithms used on large sets of Medipix data. 14 for calibrating the Timepix detectors used in LUCID. We present TAPAS, a new platform for Medipix/Timepix data that has been used extensively by secondary school Alongside the authors, the following students have also researchers in the UK for managing large amounts of contributed substantially over the history of the project: radiological data from a wide range of sources. TAPAS Toby Freeland, Matt Harrison, Cal Hewitt, Sam Kittle, permits easy storage, sharing and visualisation via a Nick Liu, Rachel O'Leary, Rachel Powell, Adam Sandey, browser interface. Hector Stalker, Tom Stevenson and Cassie Warren, as well as extensive support and encouragement over many Finally we discuss some early results from the reduced years from Dr Tom Whyntie (Oxford). Thank you also LUCID data. We nd that the majority of particles to students who contributed to the classi cations used in detected are electrons, in agreement with the trapped the training set for the machine learning classi cations, electron model and earlier simulations of LUCID. Future and the many other students who worked on LUCID over work will focus on using LUCID data to characterise the last ten years. the radiation environment of LEO (in particular using angular information), and contribute towards planned A huge personal thank you from the authors to Prof. future use of Medipix on space missions. Becky Parker (IRIS) and Laura Thomas (IRIS) who have made LUCID and CERN@school happen, and have Key results: inspired countless students across the country over many years. We present the rst results from LUCID, a novel space radiation detector that use Timepix detectors AS and EF would like to thank Mr Rupert Champion in orbit on TechDemoSat-1 (Langton Star Centre) and Dr Tim Lesworth (Langton Star Centre) for providing ongoing support with their We have developed the TAPAS platform, a versa- IRIS activities. PH acknowledges funding from the tile tool for managing Timepix datasets from a wide Engineering and Physical Sciences Research Council. range of sources. We used TAPAS to handle large amounts of LUCID and CERN@school data, lay- The authors would like to thank Prof. Steven Rose ing the groundwork for more widespread adoption (Imperial/Oxford) and Dr Carlos Granja (Advacam, of Timepix Prague) for their helpful comments, remarks and input to We present the use of machine learning algorithms the paper. to classify track morphology and thus particle type in Medipix data, allowing for quick, accurate particle The authors and the Institute for Research in classi cation for huge amounts of data Schools would like to acknowledge generous support from: Humphrey Battcock, The Science and Technology Our results add to the evidence that Medipix Facilities Council, The Science Museum, The Institute detectors are well suited as space weather moni- of Physics, The Royal Commission for the Exhibi- tors/detectors tion of 1851, The Ogden Trust, CERN, The Medipix Collaboration, NASA, the UK Space Agency, Kent We have made preliminary ux maps (showing the County Council, IEAP CTU Prague, SEEDA and the SAA and similar features) at an altitude of 600km, GridPP Collaboration (in particular Dr Dan Traynor which appear similar to those of other detectors and and Queen Mary University of London for use of their present several other early science results, including GPU and storage resources, and Professor Steve Lloyd detecting heavier charged particles for supporting and enabling schools to work with GridPP). CERN@school has acted as a large scale trial of the use of Medipix in diverse environments by sizeable The authors would also like to gratefully thank the numbers of users anonymous referees whose comments greatly improved the quality of this paper. Acknowledgements The authors and IRIS are extremely grateful to the Appendix A. Code Access Medipix Collaboration and SSTL for ten years of support, in particular we are hugely thankful to Dr Michael Camp- All the code discussed in this work is open source bell (CERN/Medipix Collaboration), Prof. Larry Pinksy and available online: https://github.com/amshenoy/ (University of Houston/NASA), David Cooke (SSTL), lucid_neural_analysis Dr Stuart Eves (SSTL) and Dr Jonathan Eastwood https://github.com/willfurnell/lucid-grid/ (Imperial). Speci c thanks to Prof. Stanislav Pospisil https://github.com/InstituteForResearchInSchools/ and IEAP CTU Prague for their support, in particular 15 lucid_utils rather than around them. If the cluster is only one pixel in size, then the density is by default set to 1. Information about LUCID, some frames from the Calculate Non-Linearity (Line Residual) - This func- mission, and some visualisation tools can be found at: tion returns the angle  anticlockwise from the x- http://starserver.researchinschools.org/lucid_ axis, with the line passing through the cluster cen- dashboard troid. First, all the coordinates of the pixels within To support and assist the production of labelled LU- the cluster are split into separate lists of X and Y CID data, the training web application can be found at: values. Single pixel tracks are by default given an an- http://starserver.researchinschools.org/lucid_ gle and line residual of 0 as these are completely lin- trainer ear. For all other clusters, the above least squares re- The TAPAS platform can be found at https://tapas. gression function is used to calculate the line of best researchinschools.org/. IRIS can be contacted for in- t in the form . From the line of best t, the best- t quiries about accessing LUCID data via TAPAS at the angle (anti-clockwise from the x-axis) of the line-of- supplied email address. best- t can be calculated using simple trigonometry. XY XY b = (X) Appendix B. Feature De nitions This appendix gives the de nitions of the features used x a = Y bX; where X = and Y = n n in the `Metric-Based Network' algorithm in Section 3, and the details of the analytic classi er. P PixelCount Line Residual = (x sin  y cos n n n=0 x sin  + y cos ) c c Find Average Neighbours - For each individual pixel in the cluster, the number of neighbouring pixels is counted by iterating through the surrounding pixel This best t angle is used to calculate the line resid- coordinates and checking if they exist in the cluster. ual value using the equation above. The line residual The pixels that exist are then stored in a list as these value is calculated by the sum of the squares of the are the ones that have an energy deposit. The mean distance from each pixel coordinate to the line travel- of the items in the list is then calculated and the ling through the centroid where the line is calculated function returns the average number of neighbours. using the best t angle . The equation above shows the simplest form of the line residual where (x , y ) c c Find Centroid (Centroid) - The centroid of the clus- is the coordinate of the centroid and (x , y ) is the n n ter is the centre of the particle track. This centroid is coordinate of the pixels in the cluster where n is the found by calculating the mean of the x values and the index of the pixel in the cluster. mean of the y values of the pixels in the cluster. The mean x-value is the presumed to be the x-coordinate Find Best Fit Circle (Curvature Radius & Circle of the centre point and the mean y-value is presumed Residual) - This function is used to perform circle to be the y-coordinate of the centre point. This co- regression. For single pixel tracks, this cannot be ordinate is not the centre of the smallest enclosing done and therefore single pixels are by default given circle but in fact, it is the centre of mass with no values of zero. As the cluster is a very poor initial weighting on energy values. This function is used to guess for the centre of the circle, multiple test calculate an auxiliary metric as the centroid is used circles are generated using the calculated best t to generate other metrics but it is not explicitly used angle from the regression line calculation. For each for classi cation as it does not have any informative of these test circles, the program performs circle value by itself. regression on the cluster using least squares. The Euclidean distance between the data points and the Calculate Radius (Radius & Diameter) - mean circle centred at (x , y ) is calculated. The p c c 2 2 Radius = (x x ) + (y y ) c n c n mean distance is then subtracted from all of the calculated distances. This value is optimised using The Euclidean distance between the coordinates of the least squares function. The optimisation process each pixel to the centroid is calculated. The radius returns an optimised centre point. This optimised is set to the highest distance from the centroid. The centre point is used to calculate the distance of all diameter is then simply twice the calculated radius. the points from the centre of the circle once again. Calculate Density (Density) - The mean distance is set as the new radius and PixelCount the residual is calculated by the summation of the Density = squares of the di erence between the distance from The density can be greater than 1 as the cluster's each point to the optimised centre point and the radius passes through the centre of the outer pixels mean radius. 16 p 2 2 (x x ) +(y y ) c n c n (Mean Radius) R = PixelCount 2 2 Circle Residual =  ( (x x ) + (y y ) c n c n n=0 R ) The optimised test circles are then organised in order of the magnitude of the residual. The aim of the optimisation is to reduce the circle residual as much as possible and therefore the circle with the least residual is chosen as the best t circle. Analytic Classi er - The analytic classi er for parti- cles used in early LUCID analysis is based on metrics that can be found in the open source lucid utils library: https: //github.com/InstituteForResearchInSchools/ lucid_utils/blob/master/lucid_utils/classification/ lucid_algorithm.py 17 REFERENCES Medipix3 pixel readout chip. Journal of Instrumentation 6 (01), C01052{C01052. URL http://stacks.iop.org/1748-0221/6/i=01/a=C01052?key= crossref.7800cb0057377fce4b91848ea8e839f1 References Beck, M. R., Scarlata, C., Fortson, L. F., Lintott, C. J., Simmons, B. D., Galloway, M. A., et al., 6 2018. Integrating human and Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, machine intelligence in galaxy morphology classi cation tasks. C., et al., 3 2016. TensorFlow: Large-Scale Machine Learning on Monthly Notices of the Royal Astronomical Society 476 (4), Heterogeneous Distributed Systems. 5516{5534. URL http://arxiv.org/abs/1603.04467 URL https://academic.oup.com/mnras/article/476/4/5516/ Accorsi, R., Celentano, L., Laccetti, P., Lanza, R. C., Marotta, M., Mettivier, G., et al., 2008. High-Resolution $^f125g$I Small Ani- Bell, A. R., Araudo, A. T., Matthews, J. H., Blundell, K. M., 1 mal Imaging With a Coded Aperture and a Hybrid Pixel Detector. 2018. Cosmic-ray acceleration by relativistic shocks: limits and IEEE Transactions on Nuclear Science 55 (1), 481{490. estimates. Monthly Notices of the Royal Astronomical Society URL http://ieeexplore.ieee.org/document/4448434/ 473 (2), 2364{2371. Ackermann, M., Ajello, M., Allafort, A., Baldini, L., Ballet, J., Bar- URL https://academic.oup.com/mnras/article/473/2/2364/ biellini, G., et al., 2 2013. Detection of the characteristic pion- decay signature in supernova remnants. Science (New York, N.Y.) Bisogni, M. G., Campbell, M., Conti, M., Delogu, P., Fantacci, 339 (6121), 807{11. M. E., Heijne, E. H. M., et al., 11 1998. Performance of a 4096- URL http://www.ncbi.nlm.nih.gov/pubmed/23413352 pixel photon counting chip&lt;/title&gt;. Vol. 3445. International Agostinelli, S., Allison, J., Amako, K., Apostolakis, J., Araujo, Society for Optics and Photonics, pp. 298{304. H., Arce, P., et al., 7 2003. Geant4a simulation toolkit. Nuclear URL http://proceedings.spiedigitallibrary.org/ Instruments and Methods in Physics Research Section A: Ac- proceeding.aspx?articleid=957789 celerators, Spectrometers, Detectors and Associated Equipment Blanchot, G., Chmeissani, M., D az, A., D az, F., Fern andez, J., 506 (3), 250{303. Garc a, E., et al., 12 2006. Dear-Mama: A photon counting X-ray URL https://www.sciencedirect.com/science/article/pii/ imaging project for medical applications. Nuclear Instruments S0168900203013688 and Methods in Physics Research Section A: Accelerators, Aguilar, M., Alberti, G., Alpat, B., Alvino, A., Ambrosi, G., Spectrometers, Detectors and Associated Equipment 569 (1), Andeen, K., et al., 4 2013. First Result from the Alpha Magnetic 136{139. Spectrometer on the International Space Station: Precision URL https://www.sciencedirect.com/science/article/pii/ Measurement of the Positron Fraction in Primary Cosmic Rays S0168900206016147 of 0.5350 GeV. Physical Review Letters 110 (14), 141102. Blasi, P., 11 2013. The origin of galactic cosmic rays. The Astronomy URL https://link.aps.org/doi/10.1103/PhysRevLett.110. and Astrophysics Review 21 (1), 70. URL http://link.springer.com/10.1007/s00159-013-0070-7 Aguilar, M., Ali Cavasonza, L., Ambrosi, G., Arruda, L., Attig, Bouchami, J., Guti errez, A., Holy, T., Houdayer, A., Jakubek, J., N., Aupetit, S., et al., jan 2018. Observation of New Properties Lebel, C., et al., 5 2011. Measurement of pattern recognition of Secondary Cosmic Rays Lithium, Beryllium, and Boron by eciency of tracks generated by ionizing radiation in a Medipix2 the Alpha Magnetic Spectrometer on the International Space device. Nuclear Instruments and Methods in Physics Research Station. Physical Review Letters 120 (2), 021101. Section A: Accelerators, Spectrometers, Detectors and Associated URL https://link.aps.org/doi/10.1103/PhysRevLett.120. Equipment 633, S187{S189. URL https://www.sciencedirect.com/science/article/pii/ Allison, J., Amako, K., Apostolakis, J., Araujo, H., Arce Dubois, P., S0168900210013537 Asai, M., et al., 2 2006. Geant4 developments and applications. Branchesi, M., 5 2016. Multi-messenger astronomy: gravitational IEEE Transactions on Nuclear Science 53 (1), 270{278. waves, neutrinos, photons, and cosmic rays. Journal of Physics: URL http://ieeexplore.ieee.org/document/1610988/ Conference Series 718 (2), 022004. Allison, J., Amako, K., Apostolakis, J., Arce, P., Asai, M., Aso, URL http://stacks.iop.org/1742-6596/718/i=2/a=022004? T., et al., 11 2016. Recent developments in Geant4. Nuclear key=crossref.f4836af5970c7a2c85257dcd576d3ecc Instruments and Methods in Physics Research Section A: Ac- Britton, D., Cass, A., Clarke, P., Coles, J., Colling, D., Doyle, celerators, Spectrometers, Detectors and Associated Equipment A., et al., 6 2009. GridPP: the UK grid for particle physics. 835, 186{225. Philosophical Transactions of the Royal Society A: Mathematical, URL https://www.sciencedirect.com/science/article/pii/ Physical and Engineering Sciences 367 (1897), 2447{2457. S0168900216306957 URL http://rsta.royalsocietypublishing.org/cgi/doi/10. Amato, E., Blasi, P., 4 2017. Cosmic ray transport in the Galaxy: 1098/rsta.2009.0036 A review. Advances in Space Research. Buncic, P., Sanchez, C. A., Blomer, J., Franco, L., Harutyunian, A., URL https://www.sciencedirect.com/science/article/pii/ Mato, P., et al., 4 2010. CernVM a virtual software appliance for S0273117717302995 LHC applications. Journal of Physics: Conference Series 219 (4), Amendolia, S., Bertolucci, E., Bisogni, M., Bottigli, U., Ceccopieri, A., Ciocci, M., et al., 2 1999. MEDIPIX: a VLSI chip for a URL http://stacks.iop.org/1742-6596/219/i=4/a=042003? GaAs pixel detector for digital radiology. Nuclear Instruments key=crossref.07e37c1115c8c3c79a47061fd61faaf8 and Methods in Physics Research Section A: Accelerators, Butzer, J. S., Butler, A. P., Butler, P., Bones, P., Cook, N., Tlustos, Spectrometers, Detectors and Associated Equipment 422 (1-3), L., 11 2008. Medipix imaging - evaluation of datasets with PCA. 201{205. In: 2008 23rd International Conference Image and Vision Com- URL https://www.sciencedirect.com/science/article/pii/ puting New Zealand. IEEE, pp. 1{6. S0168900298010948 URL http://ieeexplore.ieee.org/document/4762080/ Baca, T., Platkevic, M., Jakubek, J., Inneman, A., Stehlikova, Campbell, M., Heijne, E., Meddeler, G., Pernigotti, E., Snoeys, W., V., Urban, M., et al., 10 2016. Miniaturized X-ray telescope for 6 1998. A readout chip for a 6464 pixel matrix with 15-bit single VZLUSAT-1 nanosatellite with Timepix detector. Journal of photon counting. IEEE Transactions on Nuclear Science 45 (3), Instrumentation 11 (10), C10007{C10007. 751{753. URL http://stacks.iop.org/1748-0221/11/i=10/a=C10007? URL http://ieeexplore.ieee.org/document/682629/ key=crossref.c40da9dc58b922546561e9a2917ae846 Castellina, A., Donato, F., 2013. Astrophysics of Galactic Charged Ballabriga, R., Blaj, G., Campbell, M., Fiederle, M., Grei enberg, Cosmic Rays. In: Planets, Stars and Stellar Systems. Springer D., Heijne, E. H. M., et al., 1 2011. Characterization of the 18 Netherlands, Dordrecht, pp. 725{788. Giuliani, A., Cardillo, M., Tavani, M., Fukui, Y., Yoshiike, S., Torii, URL http://link.springer.com/10.1007/978-94-007-5612-0_ K., et al., nov 2011. Neutral pion emission from accelerated pro- 14 tons in the supernova remnant W44. The Astrophysical Journal Choi, H.-S., Lee, J., Cho, K.-S., Kwak, Y.-S., Cho, I.-H., Park, Y.- Letters, Volume 742, Issue 2, article id. L30, 5 pp. (2011). 742. D., et al., 6 2011. Analysis of GEO spacecraft anomalies: Space URL http://arxiv.org/abs/1111.4868http://dx.doi.org/10. weather relationships. Space Weather 9 (6), n/a{n/a. 1088/2041-8205/742/2/L30 URL http://doi.wiley.com/10.1029/2010SW000597 Granja, C., Jakubek, J., K oster, U., Platkevic, M., Pospisil, S., Collins, P., Akiba, K., Alexander, M., Artuso, M., Bayer, F., van 5 2011. Response of the pixel detector Timepix to heavy ions. Beuzekom, M., et al., 4 2011. The LHCb VELO upgrade. Nuclear Nuclear Instruments and Methods in Physics Research Section Instruments and Methods in Physics Research Section A: Ac- A: Accelerators, Spectrometers, Detectors and Associated Equip- celerators, Spectrometers, Detectors and Associated Equipment ment 633, S198{S202. 636 (1), S185{S193. URL https://www.sciencedirect.com/science/article/pii/ URL https://www.sciencedirect.com/science/article/pii/ S0168900210013562 S0168900210009629 Granja, C., Jakubek, J., Polansky, S., Zach, V., Krist, P., Chvatil, Colthurst, D., Kent, M., Mauchaza, R., Senasinghe, T., Sparkes, D., et al., nov 2018. Resolving power of pixel detector Timepix G., Vasilijevic, F., et al., 2015. Research science in schools: An for wide-range electron, proton and ion detection. Nuclear Instru- insight into the student experience in biomedical science and par- ments and Methods in Physics Research Section A: Accelerators, ticle physics. School Science Review 97 (358), 4451. Spectrometers, Detectors and Associated Equipment 908, 60{71. Cucinotta, F. A., 1 2007. Space Radiation Organ Doses for Astro- URL https://www.sciencedirect.com/science/article/pii/ nauts on Past and Future Missions. S0168900218309586 URL https://ntrs.nasa.gov/search.jsp?R=20070010704 Granja, C., Platkevic, M., Turecek, D., Kraus, V., Pospisil, S., Cucinotta, F. A., Manuel, F. K., Jones, J., Iszard, G., Murrey, J., Jakubek, J., et al., 2014a. Timepix-Based Miniaturized Radiation Djojonegro, B., et al., 11 2001. Space radiation and cataracts in Micro-Tracker for the Micro-Satellite RISESAT. Transactions of astronauts. Radiation research 156 (5 Pt 1), 460{6. the Japan Society for Aeronautical and Space Sciences, Aerospace URL http://www.ncbi.nlm.nih.gov/pubmed/11604058 Technology Japan 12 (ists29), Tr 7{Tr 11. D aniel, V., P na, L., Inneman, A., Zadra zil, V., B a ca, T., Platkevi c, URL http://jlc.jst.go.jp/DN/JST.JSTAGE/tastj/12.Tr_7? M., et al., 9 2016. Terrestrial gamma-ray ashes monitor demon- lang=en&from=CrossRef&type=abstract strator on CubeSat. Vol. 9978. International Society for Optics Granja, C., Polansky, S., Pospisil, S., Vykydal, Z., Turecek, D., and Photonics, p. 99780D. Owens, A., et al., 2014b. Characterization and directional visu- URL http://proceedings.spiedigitallibrary.org/ alization of space radiation quanta in low earth orbit with the proceeding.aspx?doi=10.1117/12.2240299 compact spacecraft payload SATRAM. In: Proceedings of Sci- Deligny, O., O., dec 2016. Cosmic-Ray Anisotropies: A Review. ence. eprint arXiv:1612.08002. Granja, C., Polansky, S., Vykydal, Z., Pospisil, S., Owens, A., URL http://arxiv.org/abs/1612.08002 Kozacek, Z., et al., 6 2016. The SATRAM Timepix spacecraft Denby, B., 6 1988. Neural networks and cellular automata in exper- payload in open space on board the Proba-V satellite for wide imental high energy physics. Computer Physics Communications range radiation monitoring in LEO orbit. Planetary and Space 49 (3), 429{448. Science 125, 114{129. URL https://www.sciencedirect.com/science/article/pii/ URL https://www.sciencedirect.com/science/article/pii/ 0010465588900045 S0032063316300216 Domrachev, V. V., Chugunin, D. V., 2002. Modeling of Plasma Den- Grei enberg, D., Fiederle, M., Vykydal, Z., Kr al, V., Jakubek, J., sity in the Earth's Dayside Inner Magnetosphere. Cosmic Research Holy,  T., et al., 8 2009. Detection eciency of ATLAS-MPX 40 (4), 411{417. detectors with respect to neutrons. Nuclear Instruments and URL http://link.springer.com/10.1023/A:1019806814334 Methods in Physics Research Section A: Accelerators, Spectrom- Farrell, S., Anderson, D., Cala ura, P., Cerati, G., Gray, L., eters, Detectors and Associated Equipment 607 (1), 38{40. Kowalkowski, J., et al., 8 2017. The HEP.TrkX Project: deep URL https://www.sciencedirect.com/science/article/pii/ neural networks for HL-LHC online and oine tracking. EPJ S0168900209005968 Web of Conferences 150, 00003. Greisen, K., 4 1966. End to the Cosmic-Ray Spectrum? Physical URL http://www.epj-conferences.org/10.1051/epjconf/ Review Letters 16 (17), 748{750. 201715000003 URL https://link.aps.org/doi/10.1103/PhysRevLett.16.748 Ferrari, F., Szuszkiewicz, E., may 2009. Cosmic Rays: A Review for GridPP, T. G., Faulkner, P. J. W., Lowe, L. S., Tan, C. L. A., Astrobiologists. Astrobiology 9 (4), 413{436. Watkins, P. M., Bailey, D. S., et al., 1 2006. GridPP: development URL http://www.ncbi.nlm.nih.gov/pubmed/19519216http: of the UK computing Grid for particle physics. Journal of Physics //www.liebertonline.com/doi/abs/10.1089/ast.2007.0205 G: Nuclear and Particle Physics 32 (1), N1{N20. Frisch, D. H., Smith, J. H., 5 1963. Measurement of the Relativis- URL http://stacks.iop.org/0954-3899/32/i=1/a=N01?key= tic Time Dilation Using -Mesons. American Journal of Physics crossref.9896604cdb8b56fd9e2febdf74336b83 31 (5), 342{355. Guerrini, N., Turchetta, R., Grin, D., Morse, T., Morse, A., URL http://aapt.scitation.org/doi/10.1119/1.1969508 Poyntz-Wright, O., et al., 12 2013. Design and characterisation Gaza, R., Kroupa, M., Rios, R., Stoe, N., Benton, E. R., Semones, of a highly miniaturised radiation monitor HMRM. Nuclear E. J., 11 2017. Comparison of novel active semiconductor pixel Instruments and Methods in Physics Research Section A: Ac- detectorwith passive radiation detectors during the NASA Orion celerators, Spectrometers, Detectors and Associated Equipment ExplorationFlight Test 1 (EFT-1). Radiation Measurements 106, 731, 154{159. 290{297. URL https://www.sciencedirect.com/science/article/pii/ URL https://www.sciencedirect.com/science/article/pii/ S0168900213009066?via=ihub S135044871730224X#bib22 Hartmann, B., 2013. A Novel Approach to Ion Spectroscopy of George, S. P., 2014. Mapping the Asymmetry in the SAA Fluxes Therapeutic Ion Beams Using a Pixelated Semiconductor Detec- Using the Medipix Particle-by-Particle Directional Information. tor - heiDOK. Ph.D. thesis. In: 19th WRMISS Workshop. URL http://archiv.ub.uni-heidelberg.de/volltextserver/ Ginzburg, V. L., 2 1996. Cosmic ray astrophysics (history and 15690/ general review). Physics-Uspekhi 39 (2), 155{168. Haykin, S., 1998. Neural Networks: A Comprehensive Foundation, URL http://stacks.iop.org/1063-7869/39/i=2/a=A03?key= 2nd Edition. Prentice Hall PTR, Upper Saddle River, NJ, USA. crossref.c85a1188a6f128233207236dd01fb4d4 He, K., Zhang, X., Ren, S., Sun, J., 6 2016. Deep Residual Learning 19 for Image Recognition. In: 2016 IEEE Conference on Computer Energy Cosmic Rays. Vision and Pattern Recognition (CVPR). IEEE, pp. 770{778. URL http://arxiv.org/abs/1101.4256http://dx.doi.org/10. URL http://ieeexplore.ieee.org/document/7780459/ 1146/annurev-astro-081710-102620 Heynderickx, D., Quaghebeur, B., Wera, J., Daly, E. J., Evans, H. Kroupa, M., Bahadori, A., Campbell-Ricketts, T., Empl, A., Hoang, D. R., 10 2004. New radiation environment and e ects models S. M., Idarraga-Munoz, J., et al., 7 2015. A semiconductor in the European Space Agency's Space Environment Information radiation imaging pixel detector for space radiation dosimetry. System (SPENVIS). Space Weather 2 (10), n/a{n/a. Life Sciences in Space Research 6, 69{78. URL http://doi.wiley.com/10.1029/2004SW000073 URL https://www.sciencedirect.com/science/article/pii/ Hoang, S., Pinsky, L., Vilalta, R., Jakubek, J., 12 2012. LET S2214552415000620 Estimation of Heavy Ion Particles based on a Timepix-Based Si Kroupa, M., Bahadori, A. A., Campbell-Ricketts, T., George, S. P., Detector. Journal of Physics: Conference Series 396 (2), 022023. Zeitlin, C., mar 2018. Kinetic energy reconstruction with a single URL http://stacks.iop.org/1742-6596/396/i=2/a=022023? layer particle telescope. Applied Physics Letters 112 (13), 134103. key=crossref.74ffd9c63c4ad722b5ea2caa0c667142 URL http://aip.scitation.org/doi/10.1063/1.5024920 Hoang, S. M., 2013. A PATTERN RECOGNITION APPROACH Kuwahara, T., Yoshida, K., Sakamoto, Y., Tomioka, Y., Fukuda, TO LEARNING TRACKS OF HEAVY-ION PARTICLES IN K., 12 2011. Satellite system integration based on Space Plug and TIMEPIX DETECTORS. Ph.D. thesis, University of Houston. Play Avionics. In: 2011 IEEE/SICE International Symposium on URL https://uh-ir.tdl.org/uh-ir/handle/10657/1189 System Integration (SII). IEEE, pp. 896{901. Hobbs, S., Kingston, J., Roberts, P., Juanes, C., Sewell, R., Snapir, URL http://ieeexplore.ieee.org/document/6147568/ B., et al., 1 2013. De-Orbit Sail Design for TechDemoSat-1. Li, X., Palo, S., Kohnert, R., Blum, L., Gerhardt, D., Schiller, Q., Holy, T., Heijne, E., Jakubek, J., Pospisil, S., Uher, J., Vykydal, et al., feb 2013. Small Mission Accomplished by Students-Big Im- Z., jun 2008. Pattern recognition of tracks induced by individual pact on Space Weather Research. Space Weather 11 (2), 55{56. quanta of ionizing radiation in Medipix2 silicon detector. Nuclear URL http://doi.wiley.com/10.1002/swe.20025 Instruments and Methods in Physics Research Section A: Ac- Lintott, C. J., Schawinski, K., Slosar, A., Land, K., Bamford, S., celerators, Spectrometers, Detectors and Associated Equipment Thomas, D., et al., 2008. Galaxy Zoo: morphologies derived from 591 (1), 287{290. visual inspection of galaxies from the Sloan Digital Sky Survey. URL https://www.sciencedirect.com/science/article/pii/ Monthly Notices of the Royal Astronomical Society, Volume 389, S0168900208004592 Issue 3, pp. 1179-1189. 389, 1179{1189. Jakubek, J., 5 2011. Precise energy calibration of pixel detector URL http://adsabs.harvard.edu/abs/2008MNRAS.389.1179L working in time-over-threshold mode. Nuclear Instruments and Litvak, M. L., Mitrofanov, I. G., Nuzhdin, I. O., Vostrukhin, A. V., Methods in Physics Research Section A: Accelerators, Spectrom- Golovin, D. V., Kozyrev, A. S., et al., 3 2017. Monitoring of the eters, Detectors and Associated Equipment 633, S262{S266. time and spatial distribution of neutron- ux spectral density out- URL https://www.sciencedirect.com/science/article/pii/ side the Russian segment of the International Space Station based S0168900210013732 on data from the BTN-Neutron space experiment. Cosmic Re- Jakubek, J., Cejnarova, A., Holy, T., Pospisil, S., Uher, J., Vykydal, search 55 (2), 110{123. Z., 6 2008. Pixel detectors for imaging with heavy charged URL http://link.springer.com/10.1134/S001095251702006X particles. Nuclear Instruments and Methods in Physics Research Llopart, X., Ballabriga, R., Campbell, M., Tlustos, L., Wong, Section A: Accelerators, Spectrometers, Detectors and Associated W., 10 2007. Timepix, a 65k programmable pixel readout chip Equipment 591 (1), 155{158. for arrival time, energy and/or photon counting measurements. URL https://www.sciencedirect.com/science/article/pii/ Nuclear Instruments and Methods in Physics Research Section S0168900208004245 A: Accelerators, Spectrometers, Detectors and Associated Equip- Jakubek, J., Granja, C., Hartmann, B., Jaekel, O., Martisikova, M., ment 581 (1-2), 485{494. Opalka, L., et al., 10 2011. Imaging with secondary radiation in URL http://linkinghub.elsevier.com/retrieve/pii/ hadron therapy beams with the 3D sensitive voxel detector. In: S0168900207017020 2011 IEEE Nuclear Science Symposium Conference Record. IEEE, Llopart, X., Campbell, M., Dinapoli, R., San Segundo, D., Pernig- pp. 2281{2284. otti, E., 10 2002. Medipix2: A 64-k pixel readout chip with 55-m URL http://ieeexplore.ieee.org/document/6153862/ square elements working in single photon counting mode. IEEE Jakubek, J., Granja, C., Jakel, O., Martisikova, M., Pospicil, S., Transactions on Nuclear Science 49 (5), 2279{2283. 10 2010. Detection and track visualization of primary and sec- URL http://ieeexplore.ieee.org/document/1046904/ ondary radiation in hadron therapy beams with the pixel detector Love, J. J., Pulkkinen, A., Bedrosian, P. A., Jonas, S., Kelbert, A., Timepix. In: IEEE Nuclear Science Symposuim & Medical Imag- Rigler, E. J., et al., sep 2016. Geoelectric hazard maps for the ing Conference. IEEE, pp. 1967{1969. continental United States. Geophysical Research Letters 43 (18), URL http://ieeexplore.ieee.org/document/5874118/ 9415{9424. Kataria, D. O., Fazakerley, A., Green, L., Mathews, S., Hu, H., URL http://doi.wiley.com/10.1002/2016GL070469 Hailey, M., et al., 2013. In-situ sensors and instrumentation for Marti s kov a, M., Jakubek, J., Granja, C., Hartmann, B., Op alka, L., Space Weather and the Sunjammer Solar Sail mission opportunity. Posp  sil, S., et al., 11 2011. Measurement of secondary radiation URL http://www.mssl.ucl.ac.uk/ lmg/spaceweather/Posters/ during ion beam therapy with the pixel detector Timepix. Journal SWAN_Sunjammer_Poster.pdf of Instrumentation 6 (11), C11014{C11014. Kerr, R. A., 5 2013. Planetary exploration. Radiation will make URL http://stacks.iop.org/1748-0221/6/i=11/a=C11014?key= astronauts' trip to Mars even riskier. Science (New York, N.Y.) crossref.524ac17b3fbd199ac514289b41d259a6 340 (6136), 1031. McDonnell, M. D., Tissera, M. D., Vladusich, T., van Schaik, A., URL http://www.ncbi.nlm.nih.gov/pubmed/23723213 Tapson, J., 8 2015. Fast, Simple and Accurate Handwritten Digit Khazanov, G. V., Liemohn, M. W., 5 2002. Transport of photo- Classi cation by Training Shallow Neural Network Classi ers with electrons in the nightside magnetosphere. Journal of Geophysical the Extreme Learning Machine Algorithm. PLOS ONE 10 (8), Research 107 (A5), 1064. e0134254. URL http://doi.wiley.com/10.1029/2001JA000163 URL http://dx.plos.org/10.1371/journal.pone.0134254 Komiya, Y., Shigeyama, T., 9 2017. R-process Element Cosmic Rays Michel, T., Bohnel, M., Durst, J., Sievers, P., Anton, G., 4 2009. from Neutron Star Mergers. The Astrophysical Journal 846 (2), Low Energy Dosimetry With Photon Counting Pixel Detectors 143. Such as Medipix. IEEE Transactions on Nuclear Science 56 (2), URL http://stacks.iop.org/0004-637X/846/i=2/a=143?key= 417{423. crossref.551094144e747e219a0b544ac4779f9b URL http://ieeexplore.ieee.org/document/4812304/ Kotera, K., Olinto, A. V., 1 2011. The Astrophysics of Ultrahigh Mitchell, E. F., Araujo,  H. M., Daly, E., Guerrini, N., Gunes-Lasnet, 20 S., Grin, D., et al., 7 2014. The Highly Miniaturised Radiation Bamford, S. P., Cardamone, C. N., et al., 8 2016. Galaxy Zoo: Monitor. Journal of Instrumentation 9 (07), P07010{P07010. Evidence for rapid, recent quenching within a population of AGN URL http://stacks.iop.org/1748-0221/9/i=07/a=P07010?key= host galaxies. crossref.99e925ebebf12964f35b117711b69246 URL https://arxiv.org/abs/1609.00023 Opalka, L., Granja, C., Hartmann, B., Jakubek, J., Jaekel, O., Mar- Stoe, N., Pinsky, L., Hoang, S., Idarraga, J., Kroupa, M., Jakubek, tisikova, M., et al., 2 2013. Linear energy transfer and track pat- J., et al., 12 2012. Initial results on charge and velocity discrimi- tern recognition of secondary radiation generated in hadron ther- nation for heavy ions using silicon-Timepix detectors. Journal of apy beam in a PMMA target. Journal of Instrumentation 8 (02), Instrumentation 7 (12), C12009{C12009. C02047{C02047. URL http://stacks.iop.org/1748-0221/7/i=12/a=C12009?key= URL http://stacks.iop.org/1748-0221/8/i=02/a=C02047?key= crossref.66ee75f9417ee47019dd26d97f33befe crossref.e678a45dd504d42de05d51a13b68068e Stoe, N., Pinsky, L., Kroupa, M., Hoang, S., Idarraga, J., Am- Parker, B., 2017. Real science, real classrooms. School Science Re- berboy, C., et al., 5 2015. Timepix-based radiation environment view (365). monitor measurements aboard the International Space Station. Parker, B., Fox, E., Rushton, L., 2018. IRIS extending young Nuclear Instruments and Methods in Physics Research Section peoples participation and attainment in STEM and promoting A: Accelerators, Spectrometers, Detectors and Associated Equip- teachers and students as vital members of STEM research ment 782, 143{148. communities. URL https://www.sciencedirect.com/science/article/pii/ URL https://impact.chartered.college/article/ S0168900215001977 parker-iris-stem-students-teachers-participation-research/ Taylor, B., Underwood, C., Dyer, A., Ashton, C., Rason, S., Pav on-Carrasco, F. J., De Santis, A., 4 2016. The South Atlantic Browning, J., 8 2012. The Micro Radiation Environment Mon- Anomaly: The Key for a Possible Geomagnetic Reversal. Fron- itor (MuREM) and SSTL Radiation Monitor (SSTL RM) on tiers in Earth Science 4, 40. TechDemoSat-1. IEEE Transactions on Nuclear Science 59 (4), URL http://journal.frontiersin.org/Article/10.3389/ 1060{1065. feart.2016.00040/abstract URL http://ieeexplore.ieee.org/document/6236273/ Peterson, C., 7 1989. Track nding with neural networks. Nuclear The Pierre Auger Collaboration, T. P. A., 11 2007. Correlation of Instruments and Methods in Physics Research Section A: Ac- the highest energy cosmic rays with nearby extragalactic objects. celerators, Spectrometers, Detectors and Associated Equipment URL http://arxiv.org/abs/0711.2256http://dx.doi.org/10. 279 (3), 537{545. 1126/science.1151124 URL https://www.sciencedirect.com/science/article/pii/ Thomas, S. R., Owens, M. J., Lockwood, M., 1 2014. The 22-Year 0168900289913004 Hale Cycle in Cosmic Ray Flux Evidence for Direct Heliospheric Pierre Auger Collaboration, T. P. A., Aab, A., Abreu, P., Aglietta, Modulation. Solar Physics 289 (1), 407{421. M., Samarai, I. A., Albuquerque, I. F. M., et al., 9 2017. Observa- URL http://link.springer.com/10.1007/s11207-013-0341-5 tion of a large-scale anisotropy in the arrival directions of cosmic Turecek, D., 8 2016. Arrangement of portable display system for rays above 8 1018 eV. Science (New York, N.Y.) 357 (6357), radiation detection. US Patent 9411373B2. 1266{1270. Turecek, D., Holy, T., Jakubek, J., Pospisil, S., Vykydal, Z., 1 URL http://www.ncbi.nlm.nih.gov/pubmed/28935800 2011a. Pixelman: a multi-platform data acquisition and process- Pina, L., Hudec, R., Inneman, A., Cerna, D., Jakubek, J., Sieger, ing software package for Medipix2, Timepix and Medipix3 detec- L., et al., 5 2015. X-ray monitoring for astrophysical applications tors. Journal of Instrumentation 6 (01), C01046{C01046. on Cubesat. Vol. 9510. International Society for Optics and URL http://stacks.iop.org/1748-0221/6/i=01/a=C01046?key= Photonics, p. 951005. crossref.9cbc0e538eeedde66597c91b417135cd URL http://proceedings.spiedigitallibrary.org/ Turecek, D., Pinsky, L., Jakubek, J., Vykydal, Z., Stoe, N., proceeding.aspx?doi=10.1117/12.2185091 Pospisil, S., 12 2011b. Small Dosimeter based on Timepix de- Pinsky, L., Campbell-Ricketts, T., Empl, A., George, S., Tlustos, vice for International Space Station. Journal of Instrumentation L., Turecek, D., et al., 2016. An Update on Medipix in Space 6 (12), C12037{C12037. and Future Plans (Medipix2, 3 &amp; 4). In: 6th International URL http://stacks.iop.org/1748-0221/6/i=12/a=C12037?key= Workshop on Analogue and Mixed-Signal Integrated Circuits for crossref.828425bca11aa9d1c0b4f651ad2115b9 Space Applications. Turner, D. L., Angelopoulos, V., Morley, S. K., Henderson, M. G., Pinsky, L. S., Empl, A., Gutierrez, A., Jakubek, J., Kitamura, H., Reeves, G. D., Li, W., et al., mar 2014. On the cause and extent of Miller, J., et al., may 2011. Penetrating heavy ion charge and outer radiation belt losses during the 30 September 2012 dropout velocity discrimination with a TimePix-based Si detector (for event. Journal of Geophysical Research: Space Physics 119 (3), space radiation applications). Nuclear Instruments and Methods 1530{1540. in Physics Research Section A: Accelerators, Spectrometers, URL http://doi.wiley.com/10.1002/2013JA019446 Detectors and Associated Equipment 633, S190{S193. Underwood, C., Taylor, B., Dyer, A., Hands, A., Ryden, K., Avi- URL https://www.sciencedirect.com/science/article/pii/ son, W., et al., 9 2016. Development of miniaturised radiation S0168900210013549?via{%}3Dihub environment monitors for commercial space ight: Flight results Plackett, R., Tick, T., Ballabriga, R., Vahanen, S., Turecek, D., from the MuREM/RM payloads. In: 2016 16th European Con- Campbell, M., et al., 2010. Current status of the Medipix2, ference on Radiation and Its E ects on Components and Systems Timepix, Medipix3 and Timepix2 pixel readout chips. PoS VER- (RADECS). IEEE, pp. 1{8. TEX2010, 030. URL http://ieeexplore.ieee.org/document/8093160/ URL http://inspirehep.net/record/893444?ln=en United Nations Scienti c Committee on the E ects of Atomic Radi- Potgieter, M., 2013. Solar Modulation of Cosmic Rays. Living Re- ation, 2008. SOURCES AND EFFECTS OF IONIZING RADI- views in Solar Physics 10 (1), 3. ATION UNSCEAR 2008 Report Volume I: SOURCES Report to URL http://link.springer.com/10.12942/lrsp-2013-3 the General Assembly Scienti c Annexes A and B. Tech. rep. Rawat, W., Wang, Z., 9 2017. Deep Convolutional Neural Networks Urban, M., Nentvich, O., Stehlikova, V., Baca, T., Daniel, V., for Image Classi cation: A Comprehensive Review. Neural Hudec, R., 11 2017. VZLUSAT-1: Nanosatellite with miniature Computation 29 (9), 2352{2449. lobster eye X-ray telescope and quali cation of the radiation URL http://www.mitpressjournals.org/doi/abs/10.1162/ shielding composite for space application. Acta Astronautica 140, neco_a_00990 96{104. Rossi, B., Hall, D. B., 1941. Variation of the rate of decay of URL https://www.sciencedirect.com/science/article/pii/ mesotrons with momentum. Physical Review. S0094576516308943?via%3Dihub Smethurst, R. J., Lintott, C. J., Simmons, B. D., Schawinski, K., Vilalta, R., Kuchibhotla, S., Valerio, R., Pinsky, L., 12 2011. 21 Development of Pattern Recognition Software for Tracks of Ionizing Radiation In Medipix2-Based (TimePix) Pixel Detector Devices. Journal of Physics: Conference Series 331 (3), 032052. URL http://stacks.iop.org/1742-6596/331/i=3/a=032052? key=crossref.fce3a0101cb3e64468f2cb40fdeb665c Vykydal, Z., Bouchami, J., Campbell, M., Dolezal, Z., Fiederle, M., Grei enberg, D., et al., 8 2009. The Medipix2-based network for measurement of spectral characteristics and composition of radiation in ATLAS detector. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 607 (1), 35{37. URL https://www.sciencedirect.com/science/article/pii/ S0168900209005956 Whyntie, T., Bithray, H., Cook, J., Coupe, A., Eddy, D., Fickling, R., et al., 10 2015. CERN@school: demonstrating physics with the Timepix detector. Contemporary Physics 56 (4), 451{467. URL http://www.tandfonline.com/doi/full/10.1080/ 00107514.2015.1045193 Whyntie, T., Cook, J., Coupe, A., Fickling, R. L., Parker, B., Shearer, N., 2016. CERN@school: Bringing CERN into the class- room. Nuclear and Particle Physics Proceedings. Whyntie, T., Harrison, M., 3 2015. Full simulation of the LUCID experiment in the Low Earth Orbit radiation environment. Journal of Instrumentation 10 (03), C03043{C03043. URL http://stacks.iop.org/1748-0221/10/i=03/a=C03043? key=crossref.1de52093bd3ca3bc3b66a4b5e0eee571 Whyntie, T., Harrison, M. A., 6 2014. Simulation and analysis of the LUCID experiment in the Low Earth Orbit radiation environment. Journal of Physics: Conference Series 513 (2), URL http://stacks.iop.org/1742-6596/513/i=2/a=022038? key=crossref.d670d88b21563d6f3b39d5b38202c2ac Whyntie, T., Parker, B., 5 2013. Investigating the inverse square law with the Timepix hybrid silicon pixel detector: a CERN@school demonstration experiment. Physics Education 48 (3), 344{349. URL http://stacks.iop.org/0031-9120/48/i=3/a=344?key= crossref.2fdb547cb0a30dd03c059ba251f4e8c0 Willett, K. W., Galloway, M. A., Bamford, S. P., Lintott, C. J., Masters, K. L., Scarlata, C., et al., 2 2017. Galaxy Zoo: morpho- logical classi cations for 120000 galaxies in HST legacy imaging. Monthly Notices of the Royal Astronomical Society 464 (4), 4176{4203. URL https://academic.oup.com/mnras/article-lookup/doi/ 10.1093/mnras/stw2568 Williams, D. J., Mead, G. D., 7 1965. Nightside magnetosphere con- guration as obtained from trapped electrons at 1100 kilometers. Journal of Geophysical Research 70 (13), 3017{3029. URL http://doi.wiley.com/10.1029/JZ070i013p03017 Z abori, B., Hirn, A., Deme, S., Ap athy, I., P azm andi, T., feb 2016. Characterization of cosmic rays and direction dependence in the Polar Region up to 88 km altitude. Journal of Space Weather and Space Climate 6, A12. URL http://www.swsc-journal.org/10.1051/swsc/2016006 Zeitlin, C., Hassler, D. M., Cucinotta, F. A., Ehresmann, B., Wimmer-Schweingruber, R. F., Brinza, D. E., et al., 5 2013. Mea- surements of energetic particle radiation in transit to Mars on the Mars Science Laboratory. Science (New York, N.Y.) 340 (6136), 1080{4. URL http://www.ncbi.nlm.nih.gov/pubmed/23723233

Journal

AstrophysicsarXiv (Cornell University)

Published: Oct 30, 2018

There are no references for this article.