Packora Model Systematically Predicts Molecular Crystal Structures

This paper introduces Packora, a novel flow-based generative model specifically designed for molecular crystal structure prediction (CSP). CSP is a critical task in various industries, as subtle variations in crystal structures can profoundly impact material properties. Packora distinguishes itself by jointly predicting both atomic coordinates and the crystal lattice directly from molecular graphs. Its versatility allows it to handle multi-component and organometallic crystals and to condition generation on specific molecular conformers, stereochemical labels, or space-group information. The researchers conducted a systematic study of Packora's architecture, training, conditioning, inference, and scaling, identifying key design choices for optimal performance. These include cacheable pairwise reasoning, specific training objectives, numerical solver choices, conditioning dropout, and balanced scaling of representations. Evaluated against CCDC CSP blind test standards, Packora demonstrates superior performance in both structure generation and ranking. It achieves the best coverage across generation benchmarks and shows higher recovery rates and lower ranks for experimental forms, along with faster convergence in ranking.