Performance of denoising-augmented force-matching on limited-diversity training data

Determine how the combined denoising score matching and multiscale force-matching training approach for machine-learned coarse-grained molecular dynamics force-fields performs when the available atomistic training data has limited conformational diversity, specifically whether such models can recover metastable states that are absent from the training dataset.

Background

The paper introduces a training strategy that unifies denoising score matching with bottom-up force matching to learn coarse-grained neural network force-fields from atomistic simulations. This approach aims to reduce data requirements dramatically while retaining the advantages of force-based parameterization.

In the reported experiments, the training sets—though unconverged—captured substantial conformational diversity for the proteins considered. The authors explicitly highlight uncertainty about how the proposed combined training procedure would behave when the training data is less diverse, including whether it can infer or recover states that do not appear in the training sample.