Transferable Atom-Pair Graph Neural Network for Few-Shot Reconstruction of Rigid Homobilayer Sliding Potential-Energy Surfaces

Abstract Interlayer potential-energy surfaces (PESs) govern stacking stability, sliding barriers, and tribology in two-dimensional bilayers, but first-principles mapping is expensive. We introduce an atom-pair graph neural network (APGNN) whose nodes are interlayer atom pairs, placing the representation directly on the interface. Pretraining on 89 homobilayers from a 112-material 2DMatPedia benchmark reduces the 20-point mean absolute error on 23 unseen materials from 9.22 to 1.94 meV/Å2 and increases the mean R2 from 0.802 to 0.991. Leave-one-element-out tests further demonstrate transfer to compounds containing an element absent from pretraining, reducing the MAE by 50.5–58.6% relative to that for training from scratch at the same target-data budget. For MoS2, 10 single-point energies per stacking recover the equilibrium spacing, distance-dependent sliding barriers, and two-dimensional PES topology. Dense interlayer PESs thus become accessible at a fraction of the first-principles cost. The method is currently limited to rigid homobilayers; heterobilayers and relaxed moiré structures require further training data.