Summary:

The publication explores the performance (and augmentation of) foundational models to the systems for which first principles methods become most challenging (surfaces and molecules as examples).

In this work, we identify systems for which foundational models have a different understanding of the ground state geometric configuration than DFT. For these examples, we explore both finetuning and a Gaussian process regression delta-model as improvement tactics. We demonstrate that it is possible to recover the DFT landscape with small quantities of training data, and virtually no computational investment in the case of the delta model

Furthermore, we then demonstrate that the cheap and efficient delta model can be applied to obtaining experimentally relevant theoretical predictions for very large systems in comparison to DFT. This method can be cheaply and easily applied to a wide variety of systems, providing a framework for improving the robust applicability of such potentials across materials space.

View paper here