Packora: Systematic Design for Generative Molecular Crystal Structure Prediction

Packora is an all-atom generative model designed for molecular crystal structure prediction (CSP). It addresses the challenge of predicting stable crystal packings from molecular graphs by jointly generating atomic coordinates and unit cell parameters. By supporting multi-component and organometallic systems, and allowing for flexible conditioning on molecular templates, stereochemistry, and space-group information, Packora provides a versatile framework for materials discovery.

The authors formulate CSP as a conditional generation task using variational flow matching. The model architecture is built on a Pairmixer backbone, which simplifies the complex triangle-attention mechanisms found in previous models while retaining essential pairwise reasoning. A key innovation is the post-entry injection of noisy crystal states, which allows the model's condition-only representations to be cached during the denoising process, significantly accelerating inference. The authors also introduce a two-track evaluation protocol—separating structure generation from structure ranking—to isolate the quality of the generator from the downstream relaxation and energy-ranking pipeline.

Packora demonstrates state-of-the-art performance across six standard benchmarks. In the generation track, it achieves the highest crystal-level solve rate under both standard and strict criteria, often with significant relative improvements over existing models like CLARI and OXtal. When integrated into a downstream relaxation and ranking pipeline, Packora-generated candidates lead to higher experimental-form recovery, lower energy ranks, and faster convergence during geometry optimization. The systematic study of architecture, training, and scaling reveals that balanced scaling of single and pair representations is crucial for maintaining coverage at larger model sizes.