Investigation of Data Set Portability on Various Machine-Learned Interaction Potentials for Pyrophyllite Clay

The influence of the architecture and descriptors on the accuracy of a neural network potential is investigated. Particularly, we focused on high-dimension machine-learned interaction potentials (MLIPs) using descriptors based on atom-centered symmetry functions, comprising embedding, attention masks, or message passing procedures, allowing us to include information about the atomic neighbor environment. Four potentials trained on an identical representative data set are validated, and their performances are compared against reference PBE-D3 results. All of our MLIPs, regardless of the descriptors, are able to closely reproduce PBE-D3 energy and forces. They all have similar accuracy for the different parameters investigated (structural parameters, exfoliation energy, and vibrational spectra). This shows the importance of constructing a representative data set in reaching the desired accuracy. In another effort to assess the accuracy, Grimme's D3 dispersion correction is added a posteriori to a MLIP fitted on pure PBE data. This way, the dispersion correction is not affected by the cutoff of the MLIP. This method produces accurate results for elastic parameters and vibrational spectra, showing better accuracy for exfoliation energy and great stability for structural parameters over various temperatures and cell sizes.