Modeling history-dependent behavior in granular dynamics

Develop methods that effectively model history-dependent behavior in granular dynamics, where future evolution depends on sequences of past interactions rather than solely on the current physical state.

Background

The paper identifies history dependence as a central difficulty in learned physical simulation. In granular materials, important mechanisms such as frictional tangential forces depend on the accumulated evolution of individual inter-particle contacts, including tangential displacement since contact initiation and changes in sliding direction.

Existing graph neural network simulators generally encode temporal information at the node level, for example through historical velocity frames or recurrent node states. The unresolved challenge is to represent interaction history in a way that captures the contact-scale origins of path-dependent granular behavior. TRACE addresses this challenge by introducing persistent memory states on contact edges, but the broader problem of effectively modeling history-dependent behavior remains open.

References

Despite this progress in spatial representation learning and physics-based inductive biases, effectively modeling history-dependent behavior remains an open challenge.

— TRACE: Spatiotemporal Contact Memory Graph Network Simulator for Granular Dynamics  (2609.02991 - Zhou et al., 2 Sep 2026) in Section 1, Introduction

Accurately modeling robot-foot-terrain interaction on granular media remains an open problem.

— Learning Terrain-Adaptive Humanoid Locomotion on Granular Terrain  (2609.10286 - Kamohara et al., 9 Sep 2026) in Section II-A, Granular Contact Modeling