PackFlow Generates Lattice-Aware Molecular Crystal Proposals
Researchers introduce PackFlow, a flow-matching framework for molecular crystal structure prediction, submitted Feb 23, 2026. PackFlow jointly samples heavy-atom Cartesian coordinates and unit-cell lattice parameters from molecular graphs and integrates with downstream relaxation and lattice-energy ranking. The paper also proposes physics alignment, a post-training RL stage using learned interatomic potentials to steer sampling toward low-energy basins, producing candidates that relax to lower-energy minima in blind-test evaluations.
Key Points
- 1Introduces PackFlow, a flow-matching framework generating heavy-atom coordinates and unit-cell parameters jointly
- 2Uses physics alignment reinforcement learning with learned interatomic potentials to bias generation toward low-energy regions
- 3Produces candidates that relax into lower-energy minima, improving proposal quality and overall CSP efficiency
Scoring Rationale
Novel, scalable flow-based CSP approach advances proposal generation; limited by arXiv preprint evaluation and single-source tests.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
