Simulating Complex Particle Dynamics with Graph Neural Networks

Particles are fundamental tools for simulating a wide range of physical systems, yet modeling complex particle dynamics remains a major computational challenge. This thesis explores graph neural network (GNN)-based simulators as a versatile solution to this problem, examining their effectiveness and practical limitations. Using the Graph Network-based Simulator (GNS), experiments demonstrate an ability to capture intricate behaviors across multiple domains, including water, sand, and viscoelastic materials, within a unified computational framework. Although achieving promising accuracy in short-term predictions, the model can develop instabilities during extended simulations, particularly when encountering out-of-distribution scenarios like boundary crossings. These findings highlight the potential of GNN-based methods for rapid prototyping and exploratory analysis while recognizing the need for improved robustness and scalability for broader scientific and industrial applicability.