BDIP-Net Predicts Bilayer Material Properties with Dual-Interaction Graph Learning

The discovery of new stacked bilayer materials, which possess unique properties due to the interplay of strong intra-layer and weak inter-layer forces, is computationally intensive. Traditional methods often rely on expensive DFT-based optimization for structure generation, while existing machine learning models frequently fail to differentiate between these distinct interaction types during property prediction. To address these challenges, a new machine learning framework called BDIP-Net has been developed. It combines a MatterSim-D3-based workflow for generating high-quality bilayer structures at a significantly reduced computational cost, with a novel graph neural network. BDIP-Net explicitly models both intra-layer and inter-layer interactions using specific potential representations and an adaptive message fusion mechanism. Evaluations on various bilayer systems (homobilayers, heterobilayers, twisted bilayers) demonstrate that the MatterSim-D3 workflow accurately reproduces DFT-optimized structures. Furthermore, BDIP-Net consistently surpasses other graph neural network and potential-based approaches in predicting bilayer material properties, offering a more efficient and accurate tool for materials discovery.