A comparative assessment of artificial neural network and regression models to predict mechanical properties of continuously cooled low carbon steels: an external data analysis approach

Abstract

In this study, mechanical properties of continuously cooled low carbon steels were predicted via Artificial Neural Network (ANN) and Multiple Linear Regression (MLR) models. Unlike the previous studies, laboratory scaled self-generated data that consists of chemical compositions and cooling rates were used as input while yield strength (YS), ultimate tensile strength (UTS) and total elongation (TE) were served as target data. The prediction performances of the models were compared by applying new data set extracted from external sources like previously studied research papers, thesis or dissertations. A better agreement between predicted and actual data was achieved with ANN model. Additionally, the response of ANN model to new external data resulted in lower prediction errors even the data has one or more input value that is not included in the range of training data set. Unlike ANN model, MLR model shows a significant decrease in prediction accuracy when input data has non-uniform distribution or target data takes place in relatively narrow range. In general, it was shown that ANN model trained with self-generated data can be used as an efficient tool to estimate mechanical properties of continuously cooled low carbon steels that are produced with various conditions, even for the phenomena between input and output is complex and data distribution is non-uniform.