WallpaperNet: A $p6mm$-Equivariant Graph Neural Network for Molecule Adsorption on Graphene

Rostislav Fedorov, Ganna Gryn'ova

Keywords: equivariant GNN, wallpaper-group, equivariant attention, Wyckoff-Anchor, graphene adsorption, binding-energy, hexagonal lattice, MLIP, finite group

TL;DR: WallpaperNet encodes graphene’s p6mm symmetry via Wyckoff-Anchor embeddings and D6-equivariant attention to model adsorption using only adsorbate atoms.

Abstract: We present WallpaperNet, a $p6mm$-group-equivariant graph neural network (GNN) for modeling adsorption of small molecules on graphene. Unlike conventional approaches that freeze surface atoms and operate on large system sizes, our method focuses exclusively on the adsorbate atoms, which allows fast AI-guided design of molecule-graphene composite systems. We encode the symmetry of the underlying hexagonal lattice through Wyckoff-Anchor vector encoding and $D_6$-equivariant attention mechanism.

Submission Track: Paper Track (Short Paper)

Submission Category: AI-Guided Design

Institution Location: Germany

AI4Mat RLSF: Yes

Submission Number: 109