VoroGeomNet: A Graph Neural Network Based on Voronoi Tessellation for Property Prediction of Porous Material

Porous materials have attracted extensive attention across multiple fields due to their remarkable mechanical properties. Nevertheless, predicting these properties often depends on labor-intensive experiments or computationally expensive simulations, which limit the efficiency of materials design. Recent advances in pixel-based neural networks provide a data-driven alternative. Still, they face a dilemma: high-resolution inputs impose prohibitive computational costs, whereas low-resolution inputs lead to critical information loss. To tackle this issue, we propose a resolution-agnostic graph representation based on Voronoi tessellation and introduce the Voronoi Geometric Graph Neural Network (VoroGeomNet) model with the VoroGeom Attention Convolutional Operator (VACO), specifically designed for efficient and effective message passing. Comprehensive experiments across two real-world tasks demonstrate that our model consistently outperforms the state-of-the-art baseline in prediction accuracy, achieving a maximum improvement of 51.86%, which highlights its potential for efficient and reliable screening of porous materials.