Boosting machine learning algorithms for predicting the macroscopic material behavior of continuous fiber reinforced composite

Abstract

Macroscopic mechanical properties of fibrous materials are often characterized by modeling their microscale behavior using micromechanical techniques. This process typically involves using a Representative Volume Element (RVE) and finite element simulations to obtain the macroscopic behavior through homogenization. However, these micromechanical simulations can be computationally demanding, especially for 3D models with discrete material microstructures. This paper uses boosting machine learning algorithms to predict the homogenized macroscopic material behavior of heterogeneous composites. These models are trained on the micromechanical simulation results generated by varying the constitutive parameters of local phases and microscopic parameters such as fiber volume fraction. The Bayesian optimization is used to determine the best hyperparameters of the considered boosting models, which include adaptive boosting (AdaB), gradient boosting (GBR), light gradient boosting (LGB), and extreme gradient boosting (XGB). The performances of trained models are assessed using various metrics such as R², MAE, MAPE, and RMSE and using various plots such as scatter plots, Taylor plots, radar plots, and bar plots. The comparative assessment showed that all the models predicted the homogenized stiffness matrix of the RVE successfully, with R² values between 0.94 and 0.99. The XGB model presented the best overall performance. This work contributes to the field of composites by presenting a new and computationally efficient approach to predict the macroscopic behavior of RVEs using boosting models.