Material Stiffness Prediction Based on Neural Network and Symbolic Regression

The effective stiffness of multi-inclusion composite materials is difficult to derive using theoretical methods since the localization relationships are usually complicated. In the past decades, researchers used numerical methods represented by the finite element method (FEM) to calculate the effective stiffness of multi-inclusion composites. However, FEM is time costly and cannot achieve real-time stiffness prediction. The emergence of machine learning methods provides a solution for realizing accurate real-time prediction, but making the machine learning model explainable in physical problems remains a crucial challenge. In this work, a machine learning method based on the convolutional neural network (CNN) and symbolic regression algorithm is used to achieve real-time stiffness prediction and give explicable expressions. The CNN model can predict the effective stiffness coefficients of multi-inclusion composites with all the mean absolute errors less than 2%. Using the data generated by the CNN model, the symbolic regression model based on the genetic algorithm can derive the expression of stiffness coefficients and provides a fast and explainable solution for stiffness prediction of multi-inclusion composites.