Solving Voxelized Deformable Soft Bodies with Physics-Informed Neural Networks

Abstract

Physics simulation remains a formidable challenge, particularly when balancing computational efficiency and physical fidelity. In this thesis, we investigate the feasibility of combining Physics-Informed Neural Networks (PINNs) with Interaction Networks (INs) to simulate voxelized deformable soft bodies. We begin by introducing the theoretical underpinnings of both PINNs and INs, and then detail our hybrid architecture, loss functions, and training procedures. Our experiments include a damped harmonic oscillator and a springmass deformable system, showcasing both the potential and limitations of this approach. Although the proposed method did not yield a fully stable deep learning physics simulator for large-scale voxelized deformable bodies, we discuss the factors contributing to these shortcomings and identify key failure modes. These insights, along with suggested avenues for improvement, may serve as a valuable foundation for future research in physics-based simulation and machine learning