Reproduce “Discovering Crystal Structure Prediction Algorithms with an AI Co-Scientist” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kiyoung Seong, Nayoung Kim, Sungsoo Ahn
The Human-AI Co-discovery system (HACO) developed MaskGXT, a novel masked generative transformer that discretizes and tokenizes entire crystal structures for prediction. This approach achieved state-of-the-art match rates of 73.79% on MP-20 and 36.75% on MPTS-52, alongside superior polymorph coverage with a 79.06% METRe score on the MP-20 polymorph split.
Key insights
- An AI co-scientist system can successfully identify and adapt modeling principles from diverse fields (e.g., MaskGIT from computer vision) to create new, competitive algorithms in complex scientific domains like CSP.
- Sparse, high-level human interventions, providing domain-specific constraints and objectives, effectively guide the AI agent through complex domain adaptations, leading to significant performance improvements.
- A discrete, tokenized representation of entire crystal structures, combined with masked generative learning and specific adaptations like sub-bin refinement and symmetry-aware sampling, can achieve both high geometric precision and comprehensive polymorph coverage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.