Iteratively Refined Ensemble Mlips for Aqueous Metal Oxide Molecular Dynamics

Molecular dynamics (MD) is an important method for the assessment of potential catalysts. However, classical or ab initio methods are often slow and computationally expensive. To that end, Machine Learned Interatomic Potentials (MLIPs) can be used to speed up MD simulations for easier assessment. Given that MD data is autocorrelated, we propose a sparsely sampled/coarse model that is then iteratively refined with relabeled data.

For this application, we use DeePMD-kit for the MLIPs themselves and the iterative refinement. Our models themselves are composed of a three MLIP ensemble trained on the same data. When running MD with LAMMPS, we select frames for iterative refinement based on the frame-by-frame deviation within the ensemble. For these frames, forces and potential energy surfaces are recalculated and the three models continue training with the relabeled data along with initial training data.

This procedure gives generally favorable results, with predictions for frames with unbalanced forces following a roughly 1:1 trend with DFT results. However, there are accuracy issues for force prediction when the test frame has 0 forces. Apart from this issue, the future potential of this method lies in DeePMD-kit’s agnosticism to the molecular composition of a given catalyst, potentially predicting oxide catalysts with entirely different molecular formulae from the training data with minimal training.