Predicting transient response using data-driven models for ball-impact simulations

This study investigates the application of machine learning (ML) models for predicting transient responses in ball-impact elastodynamics simulations. We focus on the canonical problem of ball impact on laminated structures, which captures essential physics while maintaining computational tractability. Novel contributions include: (1) development of a temporal multi-resolution strategy for stable long-time pre-dictions, (2) systematic comparison of U-Nets and Fourier Neural Operators as spatial ML kernels, and (3) demonstration of accurate non-local metric predictions across full time-horizons. Using a synthetic dataset of 6500 impact scenarios, we achieve 3.5-8% prediction accuracy while providing 10,000x speedup compared to traditional FEM simulations. The proposed methodology enables rapid virtual prototyping for impact-resistant design optimization.