USE OF NEURAL NETWORKS TO OPTIMIZE GRAPHITE CONTENT
IN MAGNESIA-GRAPHITE REFRACTORIES
and H. R. Savabieh
Translated from Novye Ogneupory, No. 6, pp. 41 – 46, June, 2012.
Original article submitted October 29, 2011.
Neural networks are one of the newest tools for computation. They provide unique opportunities for under
standing and studying nonlinear problems, which makes such networks well-suited for use in different fields
of engineering and technology. Only a few studies employing this new tool have been conducted in the field of
ceramics. In the study discussed in this article, a three-layer neural network was used to construct a model that
can predict the amounts of synthetic resin and graphite which graphite-bearing magnesia-carbon refractories
should contain in order to maximize their compressive strength and minimize their open porosity. The neural
network that was created can successfully predict results, and it predicted that compressive strength will be
maximal and porosity will be minimal when graphite content is within the range 10 – 17.5% and the content of
synthetic resin is 3%.
Keywords: three-layer neural network, prediction, magnesia-carbon refractories, graphite, synthetic resin,
compressive strength, porosity.
An increasing amount of interest is being shown in ex-
panding the theoretical foundation for empirically-based dy-
namic intellectual systems that can be freely modeled. An ar
tificial neural network is one such system. When used to ana
lyze production data, an artificial neural network allows rela
tionships in the data or information about it to be incorpo
rated into the network’s structure. Systems of this type are
called intellectual systems because they can be taught gen
eral rules for using this data in calculations based on images.
These systems attempt to simulate the neural-synaptic struc
ture of the human brain .
In essence, artificial neural networks (ANNs) are
data-driven black box models capable of solving complex,
highly nonlinear problems. An ANN is an information pro
cessing system that approximately replicates the behavior of
the human brain by emulating the operations and connectiv
ity of biological neurons .
Every neural network consists of neurons located in dif
ferent layers. A single neuron cannot be useful independent
of other neurons, but neurons can perform well when the are
combined with one another to form a neural network. Artifi
cial neuron models are constructed in a manner similar to
networks comprised of actual neurons. The structure of con-
temporary neural networks consists of layers made up of a
limited number of linked cells and connected to neurons in
adjacent layers .
In fact, human neurons are related to one another by
means of weight coefficients. These coefficients characterize
the strength of their connection and determine their excita
tion threshold. Thus, each neuron has an input coefficient
and an output coefficient. The behavior of networks is gener
ally determined by a transfer function. The structure of this
Refractories and Industrial Ceramics Vol. 53, No. 3, September, 2012
1083-4877/12/05303-0193 © 2012 Springer Science+Business Media New York
Iran University of Science and Technology (IUST).
Fig. 1. Diagram of a three-layer neural network : 1 ) input neu
rons; 2 ) hidden neurons; 3 ) output neurons.