Science China Information Sciences General Computer Science
Physics-informed learning for collective behavior prediction
Abstract
Collective behaviors, such as those observed in bird flocks and fish schools, exhibit highly coordinated, self-organized motion. Predicting their trajectories is crucial for understanding collective intelligence and guiding unmanned systems. Existing methods face limitations, as traditional mechanism models rely on oversimplified assumptions and generalize poorly, while purely data-driven models lack interpretability and struggle to adapt to unseen behaviors. Here, we propose a physics-informed decomposition linear neural network (PI-DLinear). By treating the physical parameters in the classic Cucker-Smale model as learnable variables, our framework implements a dynamic physics-constraining module that adapts to underlying interaction mechanisms across diverse collective behaviors. Evaluated on multiple real-world biological datasets (including jackdaw and fish), PI-DLinear significantly improves prediction accuracy within each collective behavior and consistently enhances zero-shot cross-behavior transfer relative to the architecture-matched DLinear baseline, while also demonstrating robustness in small-sample and noisy conditions. By integrating interpretable physical equations with neural networks, our work establishes a paradigm for trajectory prediction that combines interpretability with data-driven flexibility, with promising applications in biological behavior analysis and multi-agent cooperative control.