GAT-Flow: Predictor-Corrector Flow Matching with Graph Attention Network for Crystalline Materials

ICLR 2026 Conference Submission15095 Authors

Keywords: Continuous Flow Matching; Predictor-Corrector Sample Strategy; Graph Attention Network; Crystal Structure Prediction; Property-driven Materials Discovery

Abstract: Crystalline materials discovery empowered by deep generative models is critical for driving progress in applications such as energy storage, electronics, and catalysis. However, current approaches face significant challenges in accurately predicting complex structures and ensuring specific properties, thereby hindering their practical applicability. In this work, we propose GAT-Flow, a flow-based generative framework designed to address these challenges. We leverage a graph attention network to jointly predict lattice vectors and atomic coordinates, effectively capturing both local coordination and periodic patterns. We also incorporate a Predictor-Corrector sample strategy to improve sampling efficiency and numerical stability. In addition, by leveraging training-free guidance from a pre-trained language model, we enable property-driven crystalline generation based on textual prompts. Experimental results demonstrate that GAT-Flow achieves state-of-the-art performance in crystalline structure prediction. Moreover, our approach enables material generation with specific properties, offering new perspectives on structure-property alignment in computational materials design.

Primary Area: generative models

Submission Number: 15095