A Geometric Foundation Model for Crystalline Material Discovery

Shengchao Liu, Divin Yan, weitao Du, Zhuoxinran Li, Zhiling Zheng, Omar M. Yaghi, Christian Borgs, Hongyu Guo, Anima Anandkumar, Jennifer T Chayes

Submission Track: Findings & Open Challenges

Submission Category: AI-Guided Design

Keywords: foundation model, geometric modeling, geometric pretraining, property prediction, SE(3)-equivariance, periodic invariance, flow matching

TL;DR: We introduce NeuralCrystal, a geometric representation and pretraining model for crystalline material discovery.

Abstract: The use of artificial intelligence in crystalline material discovery is gaining significant attention from both the machine learning and chemistry communities. In this work, we present NeuralCrystal, a foundation model specifically designed to push the boundaries of material discovery by combining cutting-edge geometric modeling and large-scale pretraining techniques. The model ensures rotational and translational equivariance by using a vector frame basis, while projecting the coordinate system into the Fourier domain to capture the periodic symmetries and long-range interactions characteristic of crystalline materials. For geometric pretraining, we adopt an equivariant denoising approach by constructing dual views of crystalline structures from the Cambridge Structural Database. NeuralCrystal was rigorously tested on eight MatBench property prediction tasks, outperforming six, and demonstrating its strong potential to significantly accelerate the discovery of new materials.

Submission Number: 58