A comparison of machine learning methods to predict rheometric properties of rubber compounds

Abstract

In the rubber industry, rheometric properties are critical in defining processing times and temperatures. These parameters of rubber compounds are determined by time-consuming and expensive laboratory studies performed in a rheometer. Machine learning methods, on the other hand, may be used to estimate rheometric properties in seconds without the need for any samples or laboratory experiments. In this research, an artificial neural network (ANN) and two hybrid approaches of ANN with particle swarm optimisation (ANN-PSO) and genetic algorithm (ANN-GA) are used to predict the rheometric properties of a rubber compound, namely, minimum and maximum torque (ML and MH), scorch time (ts2), and 90% cure time(t90). A multi-layer perceptron (MLP) is utilised consisting of an input layer, a hidden layer, and an output layer. Whilst the network is trained by the Levenberg–Marquardt backpropagation algorithm in ANN, the network is trained by PSO and GA in hybrid approaches ANN-PSO and ANN-GA, respectively. ML, MH, ts2, and t90 are estimated using both process parameters and raw material composition as input. Dataset comprises 220 batches of a selected rubber compound. It is divided randomly into two sets as training and testing data with ratios of 85% and 15%, respectively, for each machine learning method. The prediction results are expressed as mean percentage error (MAPE). Although ANN is a powerful tool for predicting rheometric properties of rubber compounds, hybrid ANN methods decrease prediction error, resulting in better forecasts.