Discovering Crystal Structure Prediction Algorithms with an AI Co-Scientist

Researchers introduced HACO, a Human-AI co-discovery system that identified MaskGIT, a vision-based masked generative model, as an effective framework for crystal structure prediction. The resulting MaskGXT model achieved 79.06% accuracy on MP-20 benchmarks, outperforming previous baselines by 8.19 percentage points, demonstrating how AI systems can transfer learning across scientific domains when guided by human expertise.

This research demonstrates a novel paradigm where AI systems autonomously search across scientific domains to discover transferable algorithms, validated through human domain expertise. The HACO system's identification of MaskGIT—originally developed for computer vision—and its adaptation to crystallography represents a significant methodological shift in scientific discovery. Rather than humans designing algorithms from first principles, the system identified cross-domain principles and combined them with crystallographic constraints like symmetry tokens and space group stratification.

The breakthrough addresses a longstanding challenge in materials science: accurately predicting crystal structures from chemical compositions. Traditional approaches relied on physics-based simulations or domain-specific neural architectures. MaskGXT's superior performance—79.06% accuracy versus 70.87% for previous methods—suggests that masked generative modeling, proven effective in vision and NLP, generalizes well to discrete structural prediction tasks. This validates the hypothesis that scientific algorithm design benefits from systematic cross-domain exploration rather than siloed disciplinary approaches.

The practical impact extends beyond academic achievement. Materials discovery pipelines rely heavily on accurate CSP for screening candidate compounds for batteries, semiconductors, and catalysts. A 8+ percentage point accuracy improvement accelerates viable candidate identification and reduces computational screening costs. The methodology also carries broader implications: it establishes a template for AI-assisted scientific discovery in domains with fast, cheap validation—a category encompassing drug discovery, protein folding variants, and molecular optimization problems where experimental or computational validation is rapid.

Future applications may involve expanding HACO's scope to materials property prediction and extending the approach to less-validated domains where human steering becomes more critical.

• →MaskGXT achieved 79.06% accuracy on crystal structure prediction, substantially outperforming baseline methods on multiple benchmarks.
• →The HACO system successfully identified and adapted a computer vision algorithm (MaskGIT) to crystallography through cross-domain search.
• →Human-guided AI co-discovery combines autonomous algorithm exploration with targeted domain expertise rather than relying on either alone.
• →Crystal structure prediction improvements accelerate materials discovery pipelines for batteries, semiconductors, and catalysts.
• →The approach is scalable to other scientific domains with rapid validation, including drug discovery and protein design.