CoTAR: Topology and Atomic State Reconstruction in Condensed Phases

Abstract

Universal machine learning interatomic potentials (uMLIPs) enable condensed-phase molecular dynamics (MD) simulations with near-first-principles accuracy, but their lack of explicit molecular topology limits bond-aware analysis and reconnection to classical force fields. Here, we present CoTAR, a hybrid graph neural network (GNN)--hidden Markov model (HMM) framework that reconstructs molecular topology, formal charges, and unpaired electrons from atomic species, coordinates, and total charge by combining message passing on a proximity graph with a van der Waals prior, chemical constraints, and temporal smoothing. Across 128 nonreactive, topology-preserving condensed-phase systems, CoTAR achieved a bond-order-weighted F1 score of 0.906 on classical-MD data; for uMLIP trajectories, few-shot fine-tuning improved the valid-snapshot rate from 38.6\% to 84.7\%. The reconstructed topologies also supported downstream classical MD simulations, and HMM smoothing improved system-level MD simulation feasibility from 83.6\% to 85.9\%, indicating that CoTAR provides a practical framework for bond-aware analysis of condensed-phase uMLIP trajectories.