Atomistic modeling of layered polar crystals with machine learning and dispersion models

What is it about?

This paper presents an assessment of dispersion-corrected universal machine-learning interatomic potentials developed through a straightforward approach without fine-tuning or parameter refitting. These potentials are applied to investigate the crystal structure and compressibility of pnictogen chalcohalides, V–VI–VII compounds with various stoichiometries, many of which possess van der Waals gaps. The results indicate that dispersion-corrected graph deep learning potentials generally yield a more realistic description of these compounds by capturing van der Waals attractions, although this enhancement is not consistent across all cases.

Featured Image

Photo by Scott Graham on Unsplash

Photo by Scott Graham on Unsplash

Why is it important?

The development of machine-learning interatomic potentials (MLIPs) with general applicability has the potential to make atomistic materials modeling widely accessible, thanks to the computational efficiency and scalability of MLIPs. This paper explores the idea of combining a universal graph deep learning interatomic potential with generally applicable London dispersion models, addressing a crucial missing component. The studied dispersion-corrected MLIPs, in their simple and ready-to-use form, enable realistic simulations of layered polar crystals with van der Waals gaps, such as pnictogen chalcohalides, which are increasingly valued for their potential in energy applications. By facilitating realistic modeling of these materials and their composite systems, this research aims to support further theoretical studies that could drive advancements in energy technologies. Additionally, this paper introduces an automated and scalable approach for comparing computational (optimization) results with experimental characterization. This method employs two Earth mover's distances, calculated from x-ray diffraction patterns and radial distribution function histograms, to quantify the dissimilarity between the optimized structures and their corresponding experimental counterparts.