Physics-Informed Machine Learning for Crashworthiness Prediction of Hexagonal Crash Boxes under Multi-Angle Impacts

Abstract

Abstract A physics-informed machine learning framework is developed to predict temporal force–displacement responses and three-dimensional deformation histories of hexagonal crash boxes under multi-angle loading. To reduce the computational cost of nonlinear explicit finite element simulations, a three-dimensional convolutional autoencoder (3D-ConvAE) encodes crash box geometries into a low-dimensional latent representation, while a long short-term memory (LSTM) network predicts their temporal deformation evolution. A physics-informed loss function incorporating kinematic admissibility and energy conservation is introduced to suppress non-physical responses and improve mechanical consistency. Validation against benchmark numerical data demonstrates high predictive accuracy, with mean errors of 7.24% for peak force and 6.31% for energy absorption. The framework also maintains geometric fidelity across complex failure transitions, including the shift from progressive folding to global bending, achieving a mean Chamfer distance of 4 mm. The proposed approach provides a robust and computationally efficient tool for predicting nonlinear crash responses and supporting crashworthiness optimization.