Physics-Constrained Correlation-Aware Attention for Collective Cell Dynamics

Collective cell migration is shaped by short-range physical interactions between neighboring cells, but trajectory predictors often rely on heuristic attention that lacks physical grounding and interpretability. We introduce a physics-constrained, correlation-aware attention framework that embeds analytically extracted direct interaction priors into graph-based models of cell dynamics. Our approach estimates the direct correlation function from empirical structure factors using Ornstein–Zernike-based deconvolution in Fourier space and incorporates this short-range signal into attention logits as a physical prior. To encourage consistent use of this prior, we propose a variational alignment objective that regularizes learned attention distributions toward physically motivated interaction patterns via a KL divergence. This framework yields physically meaningful attention representations and provides a principled way for integrating statistical physics into representation learning for biological dynamics. We present preliminary qualitative results on collective cell migration data and outline directions for systematic evaluation and extension.