Attributing MLIP Force Predictions to Atom Groups via Model-Deviation-Gated Deletion

Machine learning interatomic potentials (MLIPs) have enabled large-scale, long-time molecular dynamics simulations. However, the force predicted for each atom combines the effects of the whole surrounding environment, and the effect of each surrounding atom or molecule is not directly obtained. This effect can be evaluated by deleting a surrounding atom group and comparing the predicted force before and after the deletion. However, if the deletion creates a local environment outside the range covered by the training data, the observed force change is not reliable. In this work we propose a post-hoc force attribution method that uses a validity gate on the deletion operation. The gate is implemented with the model deviation of an ensemble of MLIPs, and the force difference is evaluated only for the atom groups whose deletion is judged reliable. In a hydrofluoric acid solution on a silica surface, the deletions that pass the gate leave the force error between the MLIP and DFT unchanged from its value before deletion, 0.154 eV/Å, while the corresponding single-atom deletions degrade it to 0.845 eV/Å. The units obtained in this way are chemically meaningful, although no definition of molecules or bonds is supplied from outside.