Discovering Crystal Structure Prediction Algorithms with an AI Co-Scientist

The article presents the Human-AI Co-discovery system (HACO), which identifies the Masked Generative Crystal Transformer (MaskGXT) for crystal structure prediction (CSP) by leveraging generative modeling techniques from various fields. MaskGXT incorporates crystallographic symmetry tokens and space group stratified sampling, achieving a METRe accuracy of 79.06% on the MP-20 polymorph split, outperforming the closest baseline by nearly 8%. This work highlights the potential of interactive AI systems in enhancing algorithm discovery by integrating human guidance with advanced modeling strategies, offering significant implications for practitioners in materials science and computational chemistry.