New AI framework enhances crystal structure prediction accuracy

Researchers have developed CrystalGRPO, a new post-training framework for flow-based generative models aimed at improving crystal structure prediction (CSP). This framework uses reinforcement learning to optimize downstream target recovery, going beyond traditional energy rewards. CrystalGRPO integrates MACE-predicted energy with a StructureMatcher-based recovery score, offering two modes: CrystalGRPO-Q for prioritizing single-draw recovery and CrystalGRPO-C for preserving finite-budget target recovery through reference regularization and coverage-aware advantage. Experiments across various datasets and backbones demonstrated that CrystalGRPO reduces RMSE and improves Top-N recovery rates compared to existing methods. AI

IMPACT This framework could accelerate materials discovery by improving the efficiency and accuracy of predicting crystal structures.

RANK_REASON The cluster contains a research paper detailing a new AI framework for a scientific problem. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• CrystalGRPO
• Hugging Face
• MACE
• MPTS-52
• OMatG
• PXRDGen
• StructureMatcher

• paper
• model release
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →