An 8-dimensional symmetry-adapted neural network potential energy surface for H2 dissociative chemisorption on hexagonal boron nitride <i>ortho</i> site with explicit surface atom motion

Abstract

We developed an 8D, site-specific, reduced-dimensionality potential energy surface (PES) for H2 dissociation at the ortho site of hexagonal BN, treating two surface atom degrees of freedom explicitly. To our knowledge, this is the first PES for gas-phase dissociation on a covalent surface with explicit treatment of surface atom motion and symmetry in a reduced-dimensionality approach. The PES was fitted to 16 164 PBE+D3 DFT data points with a mean absolute error of 17.00 meV using a symmetry-aware neural network incorporating atom permutational invariance and lattice space group symmetry. Convergence was validated via 2D site vibrational Schrödinger equation integration and classical MD trajectory prediction uncertainty analysis. Comparison of site-specific vibrations with the surface phonon spectrum revealed the limitations of the phonon approximation for covalent PES generation using EAM-thermal averaging methods. Classical trajectory analysis reveals that H–H bond elongation beyond 1.5 Å occurs at incident energies as low as 2.25 eV, which is below the 2.664 eV chemisorption barrier, suggesting BN surface mobility facilitates pre-dissociation. Our 8D model reproduces the barrier with 17% error (0.387 eV) compared to the full 108-dimensional model, while providing at least four orders of magnitude computational speedup compared to other surface models, enabling wavepacket dynamics calculations. This methodology provides a framework for future quantum dynamics studies on covalent 2D materials, using site vibrations to construct efficient discrete-variable representation basis sets.