Learning the dynamics of particle-based systems with Lagrangian graph neural networks

Abstract Physical systems are commonly represented as a combination of particles, the individual dynamics which govern system dynamics. However, traditional approaches require knowledge several abstract quantities such energy or force to infer these particles. Here, we present framework, namely, Lagrangian graph neural network ( LGnn ), that provides strong inductive bias learn particle-based directly from trajectory. We test our approach on challenging with constraints and drag— outperforms baselines feed-forward Lnn ) improved performance. also show zero-shot generalizability by simulating two orders magnitude larger than trained one hybrid unseen model, unique feature. The architecture significantly simplifies learning in comparison ∼25 times better performance ∼20 smaller amounts data. Finally, interpretability , physical insights drag constraint forces learned model. can thus provide fillip toward understanding purely observable quantities.