CrystalGRPO Enhances Crystal Structure Prediction with Reinforcement Learning.

Flow-based generative models are effective at generating candidate crystal structures, but their initial training doesn't directly optimize for recovering specific target structures. Existing reinforcement learning (RL) methods for post-training often rely solely on energy rewards and coordinate-only policies, which can lead to a reduction in the diversity of candidates needed for successful Top-N recovery. A new framework, CrystalGRPO, addresses these limitations by aligning RL post-training with crystal structure prediction (CSP) goals. It extends policy constructions to encompass both coordinate and lattice states, integrating MACE-predicted energy with a StructureMatcher-based recovery score. CrystalGRPO offers two modes: CrystalGRPO-Q, which prioritizes single-draw recovery, and CrystalGRPO-C, designed to maintain broad candidate coverage for higher Top-N recovery. Evaluations on MP-20 and MPTS-52 datasets, using PXRDGen and OMatG backbones, demonstrate that both CrystalGRPO variants significantly reduce RMSE compared to coordinate-only RL. CrystalGRPO-Q consistently improves Top-1 recovery, while CrystalGRPO-C achieves superior Top-20 performance across all tested configurations.