Abstract
Abstract
Universal machine learning interatomic potentials (uMLIPs) enable performing condensed-phase molecular dynamics (MD) simulations with accuracy approaching that of first-principles; however, their lack of explicit molecular topology limits bond-aware analysis and reconnection to classical force fields. This study presents CoTAR, a hybrid graph neural network (GNN)–hidden Markov model (HMM) framework that reconstructs bond connectivity and order, formal charges, and unpaired electrons from atomic species and coordinates by combining learned local environments, chemical constraints, and temporal smoothing. After joint fine-tuning with 20 labeled snapshots per benchmark system, CoTAR achieved category-macro bond and conditional bond-order F1 scores of 0.992 and 0.998 and chemically valid snapshots of 82.1% across 128 uMLIP systems. Zero-shot tests were successful for all tested nonionic systems but revealed persistent formal-charge errors in ionic mixtures. For successfully reconnected systems, the reconstructed-topology simulation results showed close agreement with the reference-topology results for the density, element-resolved diffusion, and intermolecular radial distribution functions. These results establish CoTAR as a practical route from topology-free uMLIP trajectories to chemically complete molecular representations within the present scope of nonreactive, topology-preserving systems.