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

Learn More →

Efficiency testing of artificial neural networks in predicting the properties of carbon nanomaterials as potential systems for nervous tissue stimulation and regeneration

Efficiency testing of artificial neural networks in predicting the properties of carbon... A new method of predicting the properties of carbon nanomaterials from carbon nanotubes and graphene oxide, using electrophoretic deposition (EPD) on a metal surface, was investigated. The main goal is to obtain the basis for nervous tissue stimulation and regeneration. Because of the many variations of the EPD method, costly and time-consuming experiments are necessary for optimization of the produced systems. To limit such costs and workload, we propose a neural network-based model that can predict the properties of selected carbon nanomaterial systems before they are produced. The choice of neural networks as predictive learning models is based on many studies in the literature that report neural models as good interpretations of real-life processes. The use of a neural network model can reduce experimentation with unpromising methods of systems processing and preparation. Instead, it allows a focus on experiments with these systems, which are promising according to the prediction given by the neural model. The performed tests showed that the proposed method of predictive learning of carbon nanomaterial properties is easy and effective. The experiments showed that the prediction results were consistent with those obtained in the real system. Keywords: artificial neural networks; carbon nanotubes; graphene; nanomaterials; nervous tissue regeneration; stimulation. *Corresponding author: Ryszard Tadeusiewicz, Department of Control and Biomedical Engineering, AGH University of Science and Technology, al. Mickiewicza 30, 30-059 Krakow, Poland, E-mail: rtad@agh.edu.pl Martyna Sasiada: Department of Control and Biomedical Engineering, AGH University of Science and Technology, al. Mickiewicza 30, 30-059 Krakow, Poland Aneta Fraczek-Szczypta: Department of Biomaterials, AGH University of Science and Technology, al. Mickiewicza 30, 30-059 Krakow, Poland 26Sasiada et al.: Efficiency testing of artificial neural networks surface charge and surface chemistry. Experimental selection also requires the proper proportion of various types of nanomaterials (CNTs and GO) in the coating. Because all the mentioned modifications can be made regardless of each other, finding the optimal method requires many combinations. Empirical testing of all of them is very time consuming, and many reagents and real materials need to be used, including CNTs, which are expensive. It would be very useful to model and predict the above-mentioned processing by using chosen informatics tools, without carrying out numerous experiments and using real materials [21]. It is worth adding that following confirmation that this informatics tool is useful for solving a specific problem, it could also be used to solve other problems connected with different biomaterials or implants. Artificial neural network (ANN) [22] was the chosen tool in this study, as it can learn with provided examples and can also generalize the knowledge obtained [23]. It must be mentioned that using ANN in materials engineering to predict properties is a very innovative research approach. There is not much information in the literature about this kind of usage; however, it is possible to see that this subject is gaining more and more attention [24­28]. This study verified the efficiency of ANN used as a model to predict the properties of systems made of CNT and GO on a titanium surface. These systems were tested while searching for good potential materials for nervous tissue stimulation and regeneration. In order to describe the efficiency of ANN, experimental verification was made using material samples chosen by the programmed model. Samples were obtained by EPD. In this method, particles from water or organic solution were deposited on the surface of the conductive material under applied voltage. Under the influence of the electrical field, particles move either to the anode or cathode, i.e. anaphoresis or cataphoresis, respectively. Particles are deposited on the titanium surface, forming a coating [29]. Two types of multi-walled CNTs (MWCNTs) ­ commercial with -OH groups, which are described in this study as MWCNT-OH, purchased from Nanostructured & Amorphous Materials Inc. (Houston, TX, USA), and nanotubes with -OH and -COOH groups described as MWCNT-COOH, functionalized in the Department of Biomaterials at AGH University of Science and Technology ­ were used in this article to produce the coatings using the EPD method. The third type was GO from the same company (Nanostructured & Amorphous Materials Inc.). As a metal substrate for the deposition of carbon nanomaterial coatings, the titanium plate Gr2 was used. All metal samples were etched in 5% hydrofluoric acid to increase the surface development and for better adhesion of carbon coatings to the substrate. Coatings were deposited from different solutions ­ either from mixtures of acetone:ethanol:H2O (A:E:W) or from water only. Coatings from MWCNT-COOH and GO were deposited both from water and a mixture of solvents (A:E:W); however, coatings from MWCNT-OH were deposited only from A:E:W. In addition to these coatings made from MWCNT-OH, MWCNT-COOH and GO, different coatings made from CNTs and GO were prepared in this study, called hybrid coatings. They were obtained by mixing solutions of both nanomaterials in different proportions, and deposited on titanium plates using the EPD method. Coatings were applied on plates with varying voltage and time. Two types of voltage (30 and 50 V) and three different times (10, 15 and 30 s) were used. The application of deposition conditions depended on the continuity of carbon nanomaterial coatings on the titanium surface. In order to obtain learning data for ANNs, the obtained coatings were investigated using a goniometer ­ a device for contact angle measurement (automatic drop shape Experiments and results First of all, coatings made of CNT and GO were prepared before using ANN. They consisted of different proportions of the above-mentioned carbon nanomaterials. These coatings, besides having varied concentrations of CNT and GO, also had assorted types of solvent, which were used in the EPD method to apply them to the titanium surface. Selected process conditions of obtaining coatings Obtained material Examined properties Figure 1:Scheme showing each stage of experiment. Sasiada et al.: Efficiency testing of artificial neural networks27 Table 1:Results of water contact angle for all carbon nanomaterial coatings. Input data GO MWCNT-OH MWCNT-COOH Type of solution Conditions for application Voltage, V 7 50 50 50 50 50 50 50 50 50 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 50 50 50 50 50 50 50 50 50 30 30 30 30 30 30 30 30 30 Time, s 8 30 30 30 30 30 30 30 30 30 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 Output data Water contact angle, ° 9 27.6 18.3 15.4 25.3 23.1 30.1 20.3 33.5 32.5 35.5 37.2 34.5 28.9 30.4 32.4 34.2 36.1 28.8 43.7 45.9 27.8 38.9 26.1 24.2 36.7 35.7 25.3 28 32 29.4 24.2 27.8 30.1 30.5 30.8 29.3 25.1 22.9 30.3 28.7 29.4 38.8 34.8 39.5 38.5 32.8 1 C D E H J 5 Water Water Water Water Water Water Water Water Water Water Water Water Water Water Water Water Water Water ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ Water Water Water Water Water Water Water Water Water 6 ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture analysis system DSA 10Mk2; Krüss, Hamburg, Germany) and pressing diamond indenter with Berkovich geometry for obtaining the Young's modulus and hardness values. Wettability was measured using two liquids: water and diiodomethane. Knowledge of these two parameters was necessary to specify the surface energy of the obtained coatings. Mechanical parameters, which were tested, are important for the durability of coating on the platelet 28Sasiada et al.: Efficiency testing of artificial neural networks surface, because only stable coatings can be used in future experiments for nervous tissue stimulation and regeneration. Contact angle and surface energy are crucial parameters that determine the adhesion of cells to the material surface and their proliferation. The experiment scheme with carbon nanomaterials is presented in Figure 1. Table 1 presents an example listing of results for one of the tested properties (water contact angle) for all samples. The remaining parameters, like wettability with diiodomethane, wettability with water, hardness and Young's modulus, were measured in the same manner but were not presented in this study due to their size. Letters in the first column, i.e. C, D, E, H and J, are only signed ids referring to coating identification (they mean specific coatings made from nanomaterials, from a particular solution). Columns 2, 3 and 4 contain indications identifying part of individual nanomaterials in the coating. So, for the first rows, there are the numbers 0 and 1, meaning that these coatings are made from pure material; however, later, they mean different combinations. The remaining columns specify the type of solution and deposition conditions. Application and learning of ANN The main objective of the neural model used in this study is to replace real experiments (shown and described in Hypothetical technology Predicted properties Figure 2:Diagram of the ANN model predicting the experiment results. Table 2:Initial results of water contact angle used for ANN machine learning. Input data GO MWCNT-OH MWCNT-F H2O Mixture Conditions of application Normalized voltage 1 1 1 1 1 1 1 1 1 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 Normalized time 1 1 1 1 1 1 1 1 1 0.33 0.33 0.33 0.33 0.33 0.33 0.33 0.33 0.33 27.6 18.3 15.4 25.3 23.1 30.1 20.3 33.5 32.5 35.5 37.2 34.5 28.9 30.4 32.4 34.2 36.1 28.8 Output data Wettability (water), ° Reference values required during the learning on the model output 0.400000 0.095082 0.000000 0.324590 0.252459 0.481967 0.160656 0.593443 0.560656 0.659016 0.714754 0.626230 0.442623 0.491803 0.557377 0.616393 0.678689 0.439344 Sasiada et al.: Efficiency testing of artificial neural networks29 Table 3:Initial results of diiodomethane wettability used for ANN machine learning. Input data GO MWCNT-OH MWCNT-F H 2O Mixture Conditions of application Normalized voltage 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 Normalized time 1 1 1 1 1 1 1 1 0.33 0.33 0.33 0.33 0.33 0.33 0.33 0.33 Output data Wettability (diiodomethane), ° 28.1 32.6 30.7 36.6 36.3 30.6 40.5 36.8 49.3 49.3 48.6 48 45.4 43.1 46.2 46.7 Reference values required during the learning on the model output 0.000000 0.176471 0.101961 0.333333 0.321569 0.098039 0.486275 0.341176 0.831373 0.831373 0.803922 0.780392 0.678431 0.588235 0.709804 0.729412 Table 4:Initial hardness testing results used for ANN machine learning. Input data GO MWCNT-OH MWCNT-F H2O Mixture Conditions of application Normalized voltage 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Normalized time 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Output data Hardness, GPa 212 135 259 180 87 69 79 149 113 214 1093 1812 1139 2091 1643 2579 1606 2401 Reference values required during the learning on the model output 0.022713 0.010483 0.030178 0.017630 0.002859 0.000000 0.001588 0.012706 0.006989 0.023030 0.162643 0.276842 0.169949 0.321156 0.250000 0.398666 0.244123 0.370394 Figure 1). Using this model, for any hypothetical deposition condition (not yet tested), it is possible to predict the possible properties of the material created with this theoretical technology (Figure 2). The advantage of this approach is that by developing a neural model, it is possible to check many hypothetical technologies very quickly and cheaply, and for further experiments select only those materials with beneficial properties as predicted by ANN. It definitely reduces the number of experiments and allows the most profitable solution to be found at the right time. To be able to use the method shown in Figure 2, it is necessary to train an ANN capable of predicting the properties of created materials as results. Machine learning 30Sasiada et al.: Efficiency testing of artificial neural networks Table 5:Initial Young's modulus results used for ANN machine learning. Input data GO MWCNT-OH MWCNT-F H2O Mixture Conditions of application Normalized voltage 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Normalized time 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 9.7 9 1.7 6.7 6.7 3.2 3.1 5.9 4.7 9.1 86 86 95 66 73 60 91 98 Output data Young modulus, GPa Reference values required during the learning on the model output 0.045865 0.041001 0.073662 0.025017 0.025017 0.000695 0.000000 0.019458 0.011119 0.041696 0.576095 0.576095 0.638638 0.437109 0.485754 0.395413 0.610841 0.659486 Real technology Obtained material Measured properties Real technology Predicted properties Figure 3:Scheme showing one step in ANN machine learning. is a method that makes it possible to create this kind of network. To enable such learning, it is necessary to have example results from real experiments, in order to obtain nanomaterial samples (for those samples, interesting properties should be measured empirically). Table 1 lists all the results; however, if they are supposed to be used in ANN, they have to be standardized. Usually, in ANN, all values should be between [0, 1] ­ both input and output data. It is because of the mode of action of artificial neurons. This is why before working with ANN, all data were standardized by using Eq. (1): yi[0:1] = yi - ymin , ymax - ymin (1) where index i is for all the experiments conducted. Sasiada et al.: Efficiency testing of artificial neural networks31 An initial set of standardized data for the water contact angle of tested nanomaterials is presented in Table 2. Similar listings were made for the rest of the analyzed properties of carbon nanomaterials (Young's modulus, hardness and diiodomethane wettability). Primary fragments of data for specific properties are given in Tables 2­5. Table 1 already shows the big summary; this is why only primary fragments are shown in Tables 2­5. By having this data set, it is possible to now start ANN machine learning, as schematically shown in Figure 3. In one step of machine learning (which can be studied further in books that also present programs in C# for modeling and machine learning), the data of one experiment were considered, revised symbolically in Figure 3. Such an experiment delivers data about real sample processing, which was used, and real results that were obtained from material testing. An ANN receives information about the technology used as input; as the output, it should deliver predicted values of the tested properties of carbon nanomaterial coatings. At first, when ANN parameters called synaptic weights [23] have random values (because this is how they are set up at the beginning of machine learning), the result in the output obtained from ANN is far from the experimental values. This difference is called an error, and it is the factor that drives the whole machine learning process. A proper learning algorithm (e.g. backpropagation) based on the error value and also on knowledge about sample deposition conditions determines precisely which ANN parameters should be changed, and how much they should be increased or decreased. The purpose of these changes is to make the error smaller. Depending on the chosen structure, there are many parameters that are subject to tuning ­ from Figure 5:ANN structure predicting the optimal wettability of the sample using diiodomethane. Figure 6:ANN structure predicting the optimal sample hardness. Figure 4:ANN structure predicting the optimal water contact angle for the sample. Figure 7:ANN structure predicting the optimal Young's modulus of the sample. 32Sasiada et al.: Efficiency testing of artificial neural networks Table 6:ANN prediction results of water contact angle of previously unknown carbon nanomaterial systems, i.e. coatings made with previously unused technologies. Input data GO MWCNT-OH MWCNT-COOH Type of solution Conditions for application Voltage, V 50 50 50 30 30 30 50 50 50 30 30 30 50 50 50 30 30 30 50 50 50 30 30 30 50 50 50 30 30 30 50 50 50 30 30 30 Time, s 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 Prediction result received from ANN 0.435462 0.487546 0.515124 0.507030 0.613082 0.652224 0.392209 0.415466 0.429823 0.414559 0.483922 0.517977 0.405722 0.421896 0.432033 0.423577 0.475545 0.503220 0.414718 0.446509 0.464990 0.455378 0.538710 0.576141 0.454321 0.517453 0.549057 0.544768 0.656272 0.693564 0.401999 0.409583 0.414682 0.407219 0.435548 0.452362 Output data Water contact angle, ° 28.7 30.3 31.1 30.9 34.1 35.3 27.4 28.1 28.5 28.0 30.2 31.2 27.8 28.3 28.6 28.3 29.9 30.7 28.0 29.0 29.6 29.3 31.8 33.0 29.3 31.2 32.1 32.0 35.4 36.6 27.7 27.9 28.0 27.8 28.7 29.2 ­ ­ ­ ­ ­ ­ Water Water Water Water Water Water ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ Water Water Water Water Water Water ­ ­ ­ ­ ­ ­ Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture dozens to even hundreds and thousands. Thus, this detailed process is complicated, but, as a general rule, functions as it was presented here. The actions described above (in short and in simplification) are connected with one step of machine learning. On the basis of one example from the learning dataset, all subsequent examples from the same dataset are executed, and when they are completed, the learning process continues from the beginning. In effect, the whole learning dataset goes through the learning algorithm multiple times while learning. It can also occur hundreds and thousands of times until ANN is considered well learned. Used ANN structures In discussing the task of predicting the properties of carbon nanomaterial coatings, four properties are considered ­ Young's modulus, hardness, wettability with water and wettability with diiodomethane. Theoretically, it would be possible to analyze all these properties using only one neural network, which would have four outputs. However, comparative experiments showed that better results (in terms of prediction precision) were obtained while using four separate networks, each of them with one output, on which the accurate property of the system Sasiada et al.: Efficiency testing of artificial neural networks33 is analyzed. The structures of all four neural networks used are shown in Figures 4­7. These structures were created with Statistica Neural Networks from StatSoft (Statistica 2300, Tulsa, Oklahoma) by using the Neural Network Adviser module included with this suite. Input neurons mean the type of material used, mixture, and conditions of application, while output neuron means a specific material property. Colors shown in the figures stand for the values of specific neurons. The red color indicates values closest to 1, green around 0, white around 0.5 and pink means values between 0.5 and 1. Table 7:Comparison of empirical and prediction results for water contact angle. Sample ID I M P Q R Prediction result: water contact angle, ° 35.3 30.2 30.7 33 31.2 Experimental result: water contact angle, ° 50.3 39.4 28.2 41.7 36.4 Measurement error 2.71 2.01 14.3 4.66 2.66 Bold values indicate best obtained results. Table 8:Comparison of empirical and prediction results for diiodomethane wettability. Sample ID I M P Prediction result: Experimental result: Measurement diiodomethane diiodomethane error wettability, ° wettability, ° 48 48.3 44.4 49.0 50.6 46.1 2.33 2.62 3.43 Method of using ANN and results evaluation Once a properly learned neural network has been achieved, it can be used. When having some ideas for a new type of technology, it is possible to enter its parameters into ANN (especially to the above-mentioned four ANNs) and attention can be focused on predicting, by those networks, the values of the properties of tested materials, which may be created according to the analyzed technology. Most often, it shows that the predicted values are not very impressive, and even if the prediction may contain some error, usually no further experiment with using such a method is conducted. In this way, it is possible to quickly and relatively efficiently reduce the research field. Table 6 shows as an example the ANN results from water contact angle not known before carbon nanomaterial coatings, which could be created by using imagined but not empirically checked methods. Of course, for a proper evaluation of the values of predicted properties, these results obtained from the ANN output should be denormalized according to Eq. (2): yi = yi[0:1] ( ymax - ymin ) + ymin (2) Bold values indicate best obtained results. Table 9:Comparison of empirical and prediction results for hardness of coatings. Sample ID I M P Q R Prediction result: Experimental result: Measurement hardness, MPa hardness, MPa error 1130 1096 ­ 1532 327 2518 954 ­ 1194 680 1227 247 ­ 654 392 Bold values indicate best obtained results. Table 10:Comparison of empirical and prediction results for Young's modulus of coatings. Sample ID I M P Q R Prediction result: Young's modulus, GPa 50.3 45.3 ­ 45.3 32.1 Experimental result: Young's modulus, GPa 66 53 ­ 49 40 Measurement error 25 10 ­ 21 17 In a similar way, the rest of the properties of yet-unknown systems were predicted; however, relevant tables were not included in this study because of their complexity. After analysis, only five conditions were chosen, for which the predicted results were acceptable and the results were encouraging. For those five conditions, a real experiment was run, which means that the invented technology was used and new coatings were obtained, along with testing their real properties. The results of this experiment were a sort of verification ­ how much the prediction is correct compared with the results obtained empirically. Tables 7­10 show how the results for the predicted values of the four properties by ANN correspond with the Bold values indicate best obtained results. real experiment results. By analyzing these tables, it is worth mentioning that the last column is called "Measurement error". It means that shown in the "Experimental result" column are the means from 10 measurements, and their comparison with the prediction results should 34Sasiada et al.: Efficiency testing of artificial neural networks include the real measured error (standard deviation for testing). Employment or leadership: None declared. Honorarium: None declared. Competing interests: The funding organization(s) played no role in the study design; in the collection, analysis, and interpretation of data; in the writing of the report; or in the decision to submit the report for publication. Conclusions Comparison of the results received from neural networks with the results received from experiments showed that in the case of three parameters, i.e. diiodomethane wettability, Young's modulus and hardness, the results from both tests (prediction and experiments) were very similar. However, for water wettability, results from four out of the five analyzed conditions were divergent. Although not for all analyzed parameters and not for all samples, the predictions were similar. Moreover, in the range of measurement error, it is possible to say that the application of ANNs to predict the conditions of carbon nanomaterial deposition on titanium plates in the form of coating, as a potential system for nervous tissue stimulation and regeneration, is promising. In the experiments described above, ANNs allowed a significant reduction in the number of experiments that would have to be run to receive interesting information. When dealing with few material combinations that have to be tested, without using ANNs, it would be necessary to make 30 attempts of coating deposition on titanium plates. Sometimes, it would be possible to apply them properly the first time and the coating would perfectly adhere to the surface. However, sometimes, the number of attempts would have to be increased. As mentioned before, this leads not only to an increased time of conducting the experiment, but also to using more material and reagents, which are expensive, especially CNTs and graphene. Thanks to this prediction with ANN, the time for testing was shortened and the cost was reduced. Instead of 30 approaches, the number was reduced to only five. All studies will be continued concerning predicting the properties of those nanomaterials as potential systems for nerve tissue stimulation and regeneration. Author contributions: The authors have accepted responsibility for the entire content of this submitted manuscript and approved submission. Research funding: This work was supported by the National Science Centre ­ Poland (NCN; grant no. UMO2013/11/D/ST8/03272), "Carbon nanomaterial coatings on the metal surface as a potential systems for nerve cell regeneration and stimulation", financed from NCN resources. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Bio-Algorithms and Med-Systems de Gruyter

Efficiency testing of artificial neural networks in predicting the properties of carbon nanomaterials as potential systems for nervous tissue stimulation and regeneration

Download PDF
11 pages

Loading next page...
 
/lp/de-gruyter/efficiency-testing-of-artificial-neural-networks-in-predicting-the-fRr0ORCZNw
Publisher
de Gruyter
Copyright
Copyright © 2017 by the
ISSN
1895-9091
eISSN
1896-530X
DOI
10.1515/bams-2016-0025
Publisher site
See Article on Publisher Site

Abstract

A new method of predicting the properties of carbon nanomaterials from carbon nanotubes and graphene oxide, using electrophoretic deposition (EPD) on a metal surface, was investigated. The main goal is to obtain the basis for nervous tissue stimulation and regeneration. Because of the many variations of the EPD method, costly and time-consuming experiments are necessary for optimization of the produced systems. To limit such costs and workload, we propose a neural network-based model that can predict the properties of selected carbon nanomaterial systems before they are produced. The choice of neural networks as predictive learning models is based on many studies in the literature that report neural models as good interpretations of real-life processes. The use of a neural network model can reduce experimentation with unpromising methods of systems processing and preparation. Instead, it allows a focus on experiments with these systems, which are promising according to the prediction given by the neural model. The performed tests showed that the proposed method of predictive learning of carbon nanomaterial properties is easy and effective. The experiments showed that the prediction results were consistent with those obtained in the real system. Keywords: artificial neural networks; carbon nanotubes; graphene; nanomaterials; nervous tissue regeneration; stimulation. *Corresponding author: Ryszard Tadeusiewicz, Department of Control and Biomedical Engineering, AGH University of Science and Technology, al. Mickiewicza 30, 30-059 Krakow, Poland, E-mail: rtad@agh.edu.pl Martyna Sasiada: Department of Control and Biomedical Engineering, AGH University of Science and Technology, al. Mickiewicza 30, 30-059 Krakow, Poland Aneta Fraczek-Szczypta: Department of Biomaterials, AGH University of Science and Technology, al. Mickiewicza 30, 30-059 Krakow, Poland 26Sasiada et al.: Efficiency testing of artificial neural networks surface charge and surface chemistry. Experimental selection also requires the proper proportion of various types of nanomaterials (CNTs and GO) in the coating. Because all the mentioned modifications can be made regardless of each other, finding the optimal method requires many combinations. Empirical testing of all of them is very time consuming, and many reagents and real materials need to be used, including CNTs, which are expensive. It would be very useful to model and predict the above-mentioned processing by using chosen informatics tools, without carrying out numerous experiments and using real materials [21]. It is worth adding that following confirmation that this informatics tool is useful for solving a specific problem, it could also be used to solve other problems connected with different biomaterials or implants. Artificial neural network (ANN) [22] was the chosen tool in this study, as it can learn with provided examples and can also generalize the knowledge obtained [23]. It must be mentioned that using ANN in materials engineering to predict properties is a very innovative research approach. There is not much information in the literature about this kind of usage; however, it is possible to see that this subject is gaining more and more attention [24­28]. This study verified the efficiency of ANN used as a model to predict the properties of systems made of CNT and GO on a titanium surface. These systems were tested while searching for good potential materials for nervous tissue stimulation and regeneration. In order to describe the efficiency of ANN, experimental verification was made using material samples chosen by the programmed model. Samples were obtained by EPD. In this method, particles from water or organic solution were deposited on the surface of the conductive material under applied voltage. Under the influence of the electrical field, particles move either to the anode or cathode, i.e. anaphoresis or cataphoresis, respectively. Particles are deposited on the titanium surface, forming a coating [29]. Two types of multi-walled CNTs (MWCNTs) ­ commercial with -OH groups, which are described in this study as MWCNT-OH, purchased from Nanostructured & Amorphous Materials Inc. (Houston, TX, USA), and nanotubes with -OH and -COOH groups described as MWCNT-COOH, functionalized in the Department of Biomaterials at AGH University of Science and Technology ­ were used in this article to produce the coatings using the EPD method. The third type was GO from the same company (Nanostructured & Amorphous Materials Inc.). As a metal substrate for the deposition of carbon nanomaterial coatings, the titanium plate Gr2 was used. All metal samples were etched in 5% hydrofluoric acid to increase the surface development and for better adhesion of carbon coatings to the substrate. Coatings were deposited from different solutions ­ either from mixtures of acetone:ethanol:H2O (A:E:W) or from water only. Coatings from MWCNT-COOH and GO were deposited both from water and a mixture of solvents (A:E:W); however, coatings from MWCNT-OH were deposited only from A:E:W. In addition to these coatings made from MWCNT-OH, MWCNT-COOH and GO, different coatings made from CNTs and GO were prepared in this study, called hybrid coatings. They were obtained by mixing solutions of both nanomaterials in different proportions, and deposited on titanium plates using the EPD method. Coatings were applied on plates with varying voltage and time. Two types of voltage (30 and 50 V) and three different times (10, 15 and 30 s) were used. The application of deposition conditions depended on the continuity of carbon nanomaterial coatings on the titanium surface. In order to obtain learning data for ANNs, the obtained coatings were investigated using a goniometer ­ a device for contact angle measurement (automatic drop shape Experiments and results First of all, coatings made of CNT and GO were prepared before using ANN. They consisted of different proportions of the above-mentioned carbon nanomaterials. These coatings, besides having varied concentrations of CNT and GO, also had assorted types of solvent, which were used in the EPD method to apply them to the titanium surface. Selected process conditions of obtaining coatings Obtained material Examined properties Figure 1:Scheme showing each stage of experiment. Sasiada et al.: Efficiency testing of artificial neural networks27 Table 1:Results of water contact angle for all carbon nanomaterial coatings. Input data GO MWCNT-OH MWCNT-COOH Type of solution Conditions for application Voltage, V 7 50 50 50 50 50 50 50 50 50 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 50 50 50 50 50 50 50 50 50 30 30 30 30 30 30 30 30 30 Time, s 8 30 30 30 30 30 30 30 30 30 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 Output data Water contact angle, ° 9 27.6 18.3 15.4 25.3 23.1 30.1 20.3 33.5 32.5 35.5 37.2 34.5 28.9 30.4 32.4 34.2 36.1 28.8 43.7 45.9 27.8 38.9 26.1 24.2 36.7 35.7 25.3 28 32 29.4 24.2 27.8 30.1 30.5 30.8 29.3 25.1 22.9 30.3 28.7 29.4 38.8 34.8 39.5 38.5 32.8 1 C D E H J 5 Water Water Water Water Water Water Water Water Water Water Water Water Water Water Water Water Water Water ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ Water Water Water Water Water Water Water Water Water 6 ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture analysis system DSA 10Mk2; Krüss, Hamburg, Germany) and pressing diamond indenter with Berkovich geometry for obtaining the Young's modulus and hardness values. Wettability was measured using two liquids: water and diiodomethane. Knowledge of these two parameters was necessary to specify the surface energy of the obtained coatings. Mechanical parameters, which were tested, are important for the durability of coating on the platelet 28Sasiada et al.: Efficiency testing of artificial neural networks surface, because only stable coatings can be used in future experiments for nervous tissue stimulation and regeneration. Contact angle and surface energy are crucial parameters that determine the adhesion of cells to the material surface and their proliferation. The experiment scheme with carbon nanomaterials is presented in Figure 1. Table 1 presents an example listing of results for one of the tested properties (water contact angle) for all samples. The remaining parameters, like wettability with diiodomethane, wettability with water, hardness and Young's modulus, were measured in the same manner but were not presented in this study due to their size. Letters in the first column, i.e. C, D, E, H and J, are only signed ids referring to coating identification (they mean specific coatings made from nanomaterials, from a particular solution). Columns 2, 3 and 4 contain indications identifying part of individual nanomaterials in the coating. So, for the first rows, there are the numbers 0 and 1, meaning that these coatings are made from pure material; however, later, they mean different combinations. The remaining columns specify the type of solution and deposition conditions. Application and learning of ANN The main objective of the neural model used in this study is to replace real experiments (shown and described in Hypothetical technology Predicted properties Figure 2:Diagram of the ANN model predicting the experiment results. Table 2:Initial results of water contact angle used for ANN machine learning. Input data GO MWCNT-OH MWCNT-F H2O Mixture Conditions of application Normalized voltage 1 1 1 1 1 1 1 1 1 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 Normalized time 1 1 1 1 1 1 1 1 1 0.33 0.33 0.33 0.33 0.33 0.33 0.33 0.33 0.33 27.6 18.3 15.4 25.3 23.1 30.1 20.3 33.5 32.5 35.5 37.2 34.5 28.9 30.4 32.4 34.2 36.1 28.8 Output data Wettability (water), ° Reference values required during the learning on the model output 0.400000 0.095082 0.000000 0.324590 0.252459 0.481967 0.160656 0.593443 0.560656 0.659016 0.714754 0.626230 0.442623 0.491803 0.557377 0.616393 0.678689 0.439344 Sasiada et al.: Efficiency testing of artificial neural networks29 Table 3:Initial results of diiodomethane wettability used for ANN machine learning. Input data GO MWCNT-OH MWCNT-F H 2O Mixture Conditions of application Normalized voltage 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 Normalized time 1 1 1 1 1 1 1 1 0.33 0.33 0.33 0.33 0.33 0.33 0.33 0.33 Output data Wettability (diiodomethane), ° 28.1 32.6 30.7 36.6 36.3 30.6 40.5 36.8 49.3 49.3 48.6 48 45.4 43.1 46.2 46.7 Reference values required during the learning on the model output 0.000000 0.176471 0.101961 0.333333 0.321569 0.098039 0.486275 0.341176 0.831373 0.831373 0.803922 0.780392 0.678431 0.588235 0.709804 0.729412 Table 4:Initial hardness testing results used for ANN machine learning. Input data GO MWCNT-OH MWCNT-F H2O Mixture Conditions of application Normalized voltage 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Normalized time 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Output data Hardness, GPa 212 135 259 180 87 69 79 149 113 214 1093 1812 1139 2091 1643 2579 1606 2401 Reference values required during the learning on the model output 0.022713 0.010483 0.030178 0.017630 0.002859 0.000000 0.001588 0.012706 0.006989 0.023030 0.162643 0.276842 0.169949 0.321156 0.250000 0.398666 0.244123 0.370394 Figure 1). Using this model, for any hypothetical deposition condition (not yet tested), it is possible to predict the possible properties of the material created with this theoretical technology (Figure 2). The advantage of this approach is that by developing a neural model, it is possible to check many hypothetical technologies very quickly and cheaply, and for further experiments select only those materials with beneficial properties as predicted by ANN. It definitely reduces the number of experiments and allows the most profitable solution to be found at the right time. To be able to use the method shown in Figure 2, it is necessary to train an ANN capable of predicting the properties of created materials as results. Machine learning 30Sasiada et al.: Efficiency testing of artificial neural networks Table 5:Initial Young's modulus results used for ANN machine learning. Input data GO MWCNT-OH MWCNT-F H2O Mixture Conditions of application Normalized voltage 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Normalized time 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 9.7 9 1.7 6.7 6.7 3.2 3.1 5.9 4.7 9.1 86 86 95 66 73 60 91 98 Output data Young modulus, GPa Reference values required during the learning on the model output 0.045865 0.041001 0.073662 0.025017 0.025017 0.000695 0.000000 0.019458 0.011119 0.041696 0.576095 0.576095 0.638638 0.437109 0.485754 0.395413 0.610841 0.659486 Real technology Obtained material Measured properties Real technology Predicted properties Figure 3:Scheme showing one step in ANN machine learning. is a method that makes it possible to create this kind of network. To enable such learning, it is necessary to have example results from real experiments, in order to obtain nanomaterial samples (for those samples, interesting properties should be measured empirically). Table 1 lists all the results; however, if they are supposed to be used in ANN, they have to be standardized. Usually, in ANN, all values should be between [0, 1] ­ both input and output data. It is because of the mode of action of artificial neurons. This is why before working with ANN, all data were standardized by using Eq. (1): yi[0:1] = yi - ymin , ymax - ymin (1) where index i is for all the experiments conducted. Sasiada et al.: Efficiency testing of artificial neural networks31 An initial set of standardized data for the water contact angle of tested nanomaterials is presented in Table 2. Similar listings were made for the rest of the analyzed properties of carbon nanomaterials (Young's modulus, hardness and diiodomethane wettability). Primary fragments of data for specific properties are given in Tables 2­5. Table 1 already shows the big summary; this is why only primary fragments are shown in Tables 2­5. By having this data set, it is possible to now start ANN machine learning, as schematically shown in Figure 3. In one step of machine learning (which can be studied further in books that also present programs in C# for modeling and machine learning), the data of one experiment were considered, revised symbolically in Figure 3. Such an experiment delivers data about real sample processing, which was used, and real results that were obtained from material testing. An ANN receives information about the technology used as input; as the output, it should deliver predicted values of the tested properties of carbon nanomaterial coatings. At first, when ANN parameters called synaptic weights [23] have random values (because this is how they are set up at the beginning of machine learning), the result in the output obtained from ANN is far from the experimental values. This difference is called an error, and it is the factor that drives the whole machine learning process. A proper learning algorithm (e.g. backpropagation) based on the error value and also on knowledge about sample deposition conditions determines precisely which ANN parameters should be changed, and how much they should be increased or decreased. The purpose of these changes is to make the error smaller. Depending on the chosen structure, there are many parameters that are subject to tuning ­ from Figure 5:ANN structure predicting the optimal wettability of the sample using diiodomethane. Figure 6:ANN structure predicting the optimal sample hardness. Figure 4:ANN structure predicting the optimal water contact angle for the sample. Figure 7:ANN structure predicting the optimal Young's modulus of the sample. 32Sasiada et al.: Efficiency testing of artificial neural networks Table 6:ANN prediction results of water contact angle of previously unknown carbon nanomaterial systems, i.e. coatings made with previously unused technologies. Input data GO MWCNT-OH MWCNT-COOH Type of solution Conditions for application Voltage, V 50 50 50 30 30 30 50 50 50 30 30 30 50 50 50 30 30 30 50 50 50 30 30 30 50 50 50 30 30 30 50 50 50 30 30 30 Time, s 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 30 15 10 Prediction result received from ANN 0.435462 0.487546 0.515124 0.507030 0.613082 0.652224 0.392209 0.415466 0.429823 0.414559 0.483922 0.517977 0.405722 0.421896 0.432033 0.423577 0.475545 0.503220 0.414718 0.446509 0.464990 0.455378 0.538710 0.576141 0.454321 0.517453 0.549057 0.544768 0.656272 0.693564 0.401999 0.409583 0.414682 0.407219 0.435548 0.452362 Output data Water contact angle, ° 28.7 30.3 31.1 30.9 34.1 35.3 27.4 28.1 28.5 28.0 30.2 31.2 27.8 28.3 28.6 28.3 29.9 30.7 28.0 29.0 29.6 29.3 31.8 33.0 29.3 31.2 32.1 32.0 35.4 36.6 27.7 27.9 28.0 27.8 28.7 29.2 ­ ­ ­ ­ ­ ­ Water Water Water Water Water Water ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ Water Water Water Water Water Water ­ ­ ­ ­ ­ ­ Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture Mixture dozens to even hundreds and thousands. Thus, this detailed process is complicated, but, as a general rule, functions as it was presented here. The actions described above (in short and in simplification) are connected with one step of machine learning. On the basis of one example from the learning dataset, all subsequent examples from the same dataset are executed, and when they are completed, the learning process continues from the beginning. In effect, the whole learning dataset goes through the learning algorithm multiple times while learning. It can also occur hundreds and thousands of times until ANN is considered well learned. Used ANN structures In discussing the task of predicting the properties of carbon nanomaterial coatings, four properties are considered ­ Young's modulus, hardness, wettability with water and wettability with diiodomethane. Theoretically, it would be possible to analyze all these properties using only one neural network, which would have four outputs. However, comparative experiments showed that better results (in terms of prediction precision) were obtained while using four separate networks, each of them with one output, on which the accurate property of the system Sasiada et al.: Efficiency testing of artificial neural networks33 is analyzed. The structures of all four neural networks used are shown in Figures 4­7. These structures were created with Statistica Neural Networks from StatSoft (Statistica 2300, Tulsa, Oklahoma) by using the Neural Network Adviser module included with this suite. Input neurons mean the type of material used, mixture, and conditions of application, while output neuron means a specific material property. Colors shown in the figures stand for the values of specific neurons. The red color indicates values closest to 1, green around 0, white around 0.5 and pink means values between 0.5 and 1. Table 7:Comparison of empirical and prediction results for water contact angle. Sample ID I M P Q R Prediction result: water contact angle, ° 35.3 30.2 30.7 33 31.2 Experimental result: water contact angle, ° 50.3 39.4 28.2 41.7 36.4 Measurement error 2.71 2.01 14.3 4.66 2.66 Bold values indicate best obtained results. Table 8:Comparison of empirical and prediction results for diiodomethane wettability. Sample ID I M P Prediction result: Experimental result: Measurement diiodomethane diiodomethane error wettability, ° wettability, ° 48 48.3 44.4 49.0 50.6 46.1 2.33 2.62 3.43 Method of using ANN and results evaluation Once a properly learned neural network has been achieved, it can be used. When having some ideas for a new type of technology, it is possible to enter its parameters into ANN (especially to the above-mentioned four ANNs) and attention can be focused on predicting, by those networks, the values of the properties of tested materials, which may be created according to the analyzed technology. Most often, it shows that the predicted values are not very impressive, and even if the prediction may contain some error, usually no further experiment with using such a method is conducted. In this way, it is possible to quickly and relatively efficiently reduce the research field. Table 6 shows as an example the ANN results from water contact angle not known before carbon nanomaterial coatings, which could be created by using imagined but not empirically checked methods. Of course, for a proper evaluation of the values of predicted properties, these results obtained from the ANN output should be denormalized according to Eq. (2): yi = yi[0:1] ( ymax - ymin ) + ymin (2) Bold values indicate best obtained results. Table 9:Comparison of empirical and prediction results for hardness of coatings. Sample ID I M P Q R Prediction result: Experimental result: Measurement hardness, MPa hardness, MPa error 1130 1096 ­ 1532 327 2518 954 ­ 1194 680 1227 247 ­ 654 392 Bold values indicate best obtained results. Table 10:Comparison of empirical and prediction results for Young's modulus of coatings. Sample ID I M P Q R Prediction result: Young's modulus, GPa 50.3 45.3 ­ 45.3 32.1 Experimental result: Young's modulus, GPa 66 53 ­ 49 40 Measurement error 25 10 ­ 21 17 In a similar way, the rest of the properties of yet-unknown systems were predicted; however, relevant tables were not included in this study because of their complexity. After analysis, only five conditions were chosen, for which the predicted results were acceptable and the results were encouraging. For those five conditions, a real experiment was run, which means that the invented technology was used and new coatings were obtained, along with testing their real properties. The results of this experiment were a sort of verification ­ how much the prediction is correct compared with the results obtained empirically. Tables 7­10 show how the results for the predicted values of the four properties by ANN correspond with the Bold values indicate best obtained results. real experiment results. By analyzing these tables, it is worth mentioning that the last column is called "Measurement error". It means that shown in the "Experimental result" column are the means from 10 measurements, and their comparison with the prediction results should 34Sasiada et al.: Efficiency testing of artificial neural networks include the real measured error (standard deviation for testing). Employment or leadership: None declared. Honorarium: None declared. Competing interests: The funding organization(s) played no role in the study design; in the collection, analysis, and interpretation of data; in the writing of the report; or in the decision to submit the report for publication. Conclusions Comparison of the results received from neural networks with the results received from experiments showed that in the case of three parameters, i.e. diiodomethane wettability, Young's modulus and hardness, the results from both tests (prediction and experiments) were very similar. However, for water wettability, results from four out of the five analyzed conditions were divergent. Although not for all analyzed parameters and not for all samples, the predictions were similar. Moreover, in the range of measurement error, it is possible to say that the application of ANNs to predict the conditions of carbon nanomaterial deposition on titanium plates in the form of coating, as a potential system for nervous tissue stimulation and regeneration, is promising. In the experiments described above, ANNs allowed a significant reduction in the number of experiments that would have to be run to receive interesting information. When dealing with few material combinations that have to be tested, without using ANNs, it would be necessary to make 30 attempts of coating deposition on titanium plates. Sometimes, it would be possible to apply them properly the first time and the coating would perfectly adhere to the surface. However, sometimes, the number of attempts would have to be increased. As mentioned before, this leads not only to an increased time of conducting the experiment, but also to using more material and reagents, which are expensive, especially CNTs and graphene. Thanks to this prediction with ANN, the time for testing was shortened and the cost was reduced. Instead of 30 approaches, the number was reduced to only five. All studies will be continued concerning predicting the properties of those nanomaterials as potential systems for nerve tissue stimulation and regeneration. Author contributions: The authors have accepted responsibility for the entire content of this submitted manuscript and approved submission. Research funding: This work was supported by the National Science Centre ­ Poland (NCN; grant no. UMO2013/11/D/ST8/03272), "Carbon nanomaterial coatings on the metal surface as a potential systems for nerve cell regeneration and stimulation", financed from NCN resources.

Journal

Bio-Algorithms and Med-Systems – de Gruyter

Published: Mar 1, 2017

References

  • Artificial intelligence as a tool to support the development and testing of biomaterials
  • Carbon nanotube-based coatings on titanium
  • From nano to giant? Designing carbon nanotubes for rubber reinforcement and their applications for high performance tires
  • Prediction of graphite nanofluids’ dynamic viscosity by means of artificial neural networks
  • Electrical properties and applications of carbon nanotube structures
  • Carbon nanomaterials for nerve tissue stimulation and regeneration
  • Prediction of the physicochemical properties of woody biomass using linear prediction and artificial neural networks
  • Carbon nanomaterials for nerve tissue stimulation and regeneration
  • Thermal, mechanical, and electrical properties of graphene nanoplatelet/graphene oxide/polyurethane hybrid nanocomposite
  • Carbon nanotube-based coatings on titanium
  • Aligned PLLA nanofibrous scaffolds coated with graphene oxide for promoting neural cell growth