A Latent Simulator for Interacting Particle Systems

To model interacting dynamical systems accurately, it is essential to capture both spatial interactions and long-range temporal dependencies simultaneously. However, previous works consider these separately, leading to errors.

To address this problem, Karn Tiwari, Niladri Dutta, and Prathosh A P (from the Indian Institute of Science) and N M Anoop Krishnan (from the Yardi School of Artificial Intelligence) developed the Graph Mamba Operator (GraMO), a latent-space operator that integrates state–space models with graph-based interaction learning.

Figure: Overview. The model maps past graph sequences {Gt}T to future trajectories {Gt}T +∆T
using L stacked GraMO blocks. (Left) Overall pipeline. (Right) A single GraMO layer showing the discretized update and gated skip connection.

For particle-based systems, the choice of latent representation has two requirements: (i) the representation must reflect the graph structure of interactions, since the dynamics of each body depends on its neighbours through the interaction potential; and (ii) the representation must persist over time, since particle trajectories carry long-range dependencies through inertia, recurring contacts, and slow collective models.

GraMO provides a principled framework for learning complex spatiotemporal dynamics.

  • The operator evolves a persistent latent state under input-dependent dynamics, jointly modelling spatial interactions and temporal memory in a single latent recurrence.
  • The method achieves improved trajectory prediction and zero-shot generalisation across N-body simulation, motion capture, and robotics benchmarks, outperforming graph-based and neural operator baselines.

This research work has been accepted for presentation at the 40th Annual Conference on Neural Information Processing Systems (NeurIPS 2026).

Paper: https://arxiv.org/abs/2606.09432

MENU