Transferability of machine learning potentials: Protonated water neural network potential applied to the protonated water hexamer

TL;DR

This study demonstrates that a neural network potential trained on smaller protonated water clusters can accurately predict the properties of a larger, unseen cluster in an extrapolation regime, indicating strong transferability.

Contribution

The paper shows that a neural network potential trained on smaller clusters can reliably extrapolate to larger clusters, expanding the applicability of machine learning potentials beyond their training data.

Findings

The neural network potential accurately describes the protonated water hexamer.

The model maintains good accuracy across a temperature range from 1 K to 300 K.

Transferability is explained by the similarity of atomic environments in training and test systems.

Abstract

A previously published neural network potential for the description of protonated water clusters up to the protonated water tetramer, H$^+$(H$_2$O)$_4$, at essentially converged coupled cluster accuracy (J. Chem. Theory Comput. 16, 88 (2020)) is applied to the protonated water hexamer, H$^+$(H$_2$O)$_6$ -- a system that the neural network has never seen before. Although being in the extrapolation regime, it is shown that the potential not only allows for quantum simulations from ultra-low temperatures $\sim$ 1 K up to 300 K, but that it is able to describe the new system very accurately compared to explicit coupled cluster calculations. This transferability of the model is rationalized by the similarity of the atomic environments encountered for the larger cluster compared to the environments in the training set of the model. Compared to the interpolation regime the quality of the model is reduced by roughly one order of magnitude, but most of the difference to the coupled cluster reference comes from global shifts of the potential energy surface, while local energy fluctuations are well recovered. These results suggest that the application of neural network potentials in extrapolation regimes can provide useful results and might be more general than usually thought.