On the machine learning approach for computational homogenization of short fiber-reinforced composites with elastoplastic behavior

Abstract

This study aims to develop machine learning (ML)-based homogenization models to efficiently predict the effective elastoplastic properties of short fiber-reinforced composites (SFRCs), reducing the reliance on computationally expensive micromechanical simulations. Sobol sampling is employed to generate training data by varying constitutive and microstructural parameters of representative volume element. Several ML models including artificial neural networks (ANN), support vector regression (SVR), random forest (RF), and extreme gradient boosting (XGB) are trained to predict homogenized stress-strain responses. A new approach is introduced that decomposes the stress-strain response into elastic and plastic components, allowing the ML models to learn these components separately and effectively. Additionally, the Taguchi method (L27 orthogonal array) is employed to minimize simulation runs and evaluate parameter sensitivity through ANOVA. The best-performing ML model is implemented in a finite element analysis (FEA) of an automotive component. Among all models, ANN demonstrated the highest accuracy in predicting the macroscopic elastoplastic response across a wide range of input parameters. Finally, the ANN-based elastoplastic material model is validated on a real-world macroscopic structure by implementing it into the finite element analysis of an automotive component. The results demonstrate that ML-based homogenization closely matches the traditional homogenization methods and also highlights its ability to effectively capture nonlinear material behavior.