Physics-regularized neural networks learn collective motion from trajectory data

A new neural network architecture that embeds physical structure directly into its design can learn the dynamics of collective motion from trajectory data while avoiding a persistent failure mode of data-driven models: the accumulation of small errors into badly distorted long-term predictions. The approach, called a Physics-Regularized Graph Neural Network (PR-GNN), is described in an open-access study published in Complex & Intelligent Systems.

The problem the researchers targeted arises when models trained only on short-horizon transitions are deployed in closed loop, meaning each prediction is fed back as input for the next step. Small inaccuracies compound over repeated rollouts, degrading the model’s ability to reproduce the large-scale organization of a moving group even when individual short-term predictions look accurate. PR-GNN counters this by incorporating physical structure at several levels. It constructs focal-centered dynamic interaction graphs around each moving individual and estimates the importance of neighbors through a learned attention mechanism. The predicted acceleration is then organized around three physically motivated components derived from motion features: self-motion, velocity alignment, and spatial interaction.

Training is likewise shifted toward long-horizon fidelity. The authors use a multi-step rollout training objective, augmented by a collective behavior regularization term, so that learning favors stable closed-loop behavior rather than short-horizon accuracy alone. On six self-propelled-particle datasets, PR-GNN achieved lower closed-loop position errors than all five evaluated baseline methods. Crucially, the gains did not come at the expense of group-level structure: the model preserved key macroscopic indicators, including polarization, rotational organization, mean speed, and spatial spread.

Beyond predictive performance, the architecture offers interpretability. The learned attention distributions and response-channel signatures provide post-hoc evidence of behaviorally relevant neighbor relations and of the dominant physical response modes governing the motion. By combining graph-based interaction modeling, attention over neighbors, and physics-motivated decomposition of acceleration, the study suggests a route to trajectory models that remain stable under recursive deployment while keeping the macroscopic organization of collective behavior intact.

Subject of Research: A physics-regularized graph neural network that learns collective motion dynamics from trajectory data

Article Title: Learning collective motion dynamics from trajectory data with physics-regularized graph neural networks

Article References: Liu, T., Li, Y., Wang, T., & Peng, X. (2026). Learning collective motion dynamics from trajectory data with physics-regularized graph neural networks. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02566-z

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02566-z

Keywords: collective motion, graph neural networks, trajectory data, self-propelled particles, physics regularization, closed-loop rollout, attention mechanism, polarization, machine learning, complex systems, Learning, collective