Deep learning for predicting the evolution of chiral active particles motion based on experimental platform

Chiral Active Particles (CAPs) exhibit intricate collective behaviors, yet their characterization via traditional simulations or experiments is often labor-intensive and resource-demanding. To overcome these challenges, we introduce an integrated experimental platform and deep learning (DL) framework designed to predict CAP collective motion dynamics within confining structures. Our platform efficiently captures extensive spatio-temporal data of CAP assemblies under controlled vibration. The predictive model uniquely leverages Particle Density Maps (PDDMs) derived from experimental video-a crucial representation that effectively mitigates extraneous RGB image interference and enhances robustness against visual noise, thereby enabling the model to directly capture and learn collective states. The DL architecture is trained to generate the temporal evolution of these density maps, successfully capturing complex emergent phenomena like chiral separation. Rigorous validation demonstrates the framework's effectiveness, achieving a low average relative error of 8.94\% for the predicted Minimum Enclosing Circle (MEC) radius of particle clusters. It demonstrates strong generalizability across time (15 future frames from 5 inputs) and novel geometries (non-concentric annuli). This data-driven approach provides significant computational efficiency over traditional simulations (e.g., Active Brownian Particle models, which struggle with such complexities), streamlining analysis and advancing understanding of active matter systems.