Combining graph deep learning and London dispersion interatomic potentials: A case study on pnictogen chalcohalides

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Machine-learning interatomic potentials, often based on graph neural networks (GNNs), offer efficient atomistic materials modeling.
• Current GNN potentials trained on density-functional theory (DFT) data frequently lack long-range interactions like London dispersion forces.
• This omission limits their accuracy for materials where van der Waals interactions are significant.

Purpose of the Study:

• To investigate the impact of incorporating semiempirical dispersion models into GNN potentials.
• To assess if this combination improves the accuracy of predicting crystal structures and properties for layered materials.
• To evaluate the effectiveness of dispersion-corrected GNN potentials for V-VI-VII compounds.

Main Methods:

• Developed and applied dispersion-corrected GNN potentials.
• Derived equations of state for BiTeBr and BiTeI.
• Performed crystal structure optimizations for various V-VI-VII compounds.
• Characterized structures using X-ray diffraction patterns and radial distribution functions.
• Quantified structural differences using Earth Mover's Distance.

Main Results:

• Dispersion-corrected GNN potentials generally provide more realistic descriptions of the studied compounds.
• Inclusion of van der Waals attractions led to systematic improvements in predicting van der Waals gaps and layer thicknesses.
• Optimized crystal structures showed better agreement with experimental data when dispersion was included.

Conclusions:

• Combining GNN potentials with semiempirical dispersion models is a straightforward and effective approach.
• This method significantly enhances the description of layered polar crystals without requiring retraining or parameter refitting.
• Dispersion-corrected GNN potentials represent a valuable advancement for accurate atomistic materials modeling.