Molecular mechanics force-fields are essential for molecular dynamics simulations as they provide computational efficiency, stability, and physically interpretable models. However, their reliance on discrete atom types restricts transferability to chemically diverse or modified molecules. Machine-learning offers a route to overcome these limitations by learning continuous chemical representations directly from molecular graphs. We present CHARMM2GAT, a graph-attentional transformer trained to reproduce CHARMM-style bonded and non-bonded parameters solely from molecular connectivity. This framework probes the extent to which a machine can internalize the information encoded in a traditional force field, while also exposing where discretized atom-typing schemes create inconsistencies that may be resolved through targeted refinement or quantum mechanical reference data. Our results suggest that while a neural network can recover broad chemical trends from CHARMM, the discontinuities introduced by atom typing remain evident. Machine-learning thus emerges not only as a tool to automate parameter assignment but also to point out limitations of classical force fields.
Hungerland et al. (Sun,) studied this question.