Optimal Design of Nanocomposite Material and Fabrication Parameters Using Artificial Neural Network

Abstract

Artificial neural network (ANN) is a mathematical model that maps an ndimensional real input to an m-dimensional output. The basic computing unit, called a node, calculates the weighted sum of the inputs and compares it with a threshold (or bias) and sends a signal to the output after passing it through an activation function. Activation function is the mathematical function that defines the way an incoming signal at a node is treated before being outputted; commonly used functions are linear, tan sigmoid and log sigmoid. The set of nodes that receives the input from the data set and those, which deliver the final predicted outputs are called the input and output layers, respectively. Other node sets, the hidden layers, have no direct connection to the outside but are connected to the input, output and other hidden layers. An ANN model is trained, that is, the weights connecting the various nodes adjusted through an iterative process. We have use an ANN-based model to predict the flexural strength of composites. Experimental data from 96 experiments were used to train and validate the ANN. Four parameters corresponding to different manufacturing quantities have been used as the input parameter, and the two parameters representing failure load and displacement has been chosen at the output of the ANN model. The Backpropagation was used to train the ANN model, while the Levenberg-Marquardt was employed for minimizing the errors. The activation functions used are linear for the nodes in the input layer and tan sigmoid for the nodes in the hidden and output layer. We also found that the amount of data which is needed by the ANN model to obtain similar accuracy results to that of the traditional regression approach is about 40-50% less. This means we should be able to accelerate the process of composite design by almost 2 times. Also it was found that the ANN model predicted similar results, but with higher correlation coefficients and R2 values compared to the values available for reference. Additionally, we found that the model needed only 50% of the original data to produce results with similar correlation coefficients and R2 values as the values available for reference.