Where Should Physics Enter a Molecular Crystal Generator?
Abstract
Generative models make molecular crystal structure prediction fast, but their samples still exhibit geometric and packing violations. Physics can be introduced during training, post-training, or inference, yet these choices are rarely compared with the generator and physical signal held fixed. We introduce CrystAF, an all-atom crystal flow-map generation model, and use it with the UMA interatomic potential to systematically study where physics should enter. Post-training learns physical preferences directly into CrystAF, improving molecular validity and crystal packing while leaving sampling unchanged: physics is paid for once during training rather than repeatedly at deployment. In contrast, UMA relaxation is effective at repairing local clashes but makes generation 6--26× slower, while learning from relaxed targets provides little benefit. These routes are complementary rather than competing. Physics-informed post-training first shifts the generated distribution toward more physically reasonable structures, after which inexpensive inference-time corrections further remove clashes and restore stereochemistry that the generator cannot represent. Importantly, the same post-training strategy also improves the multi-step all-atom Clari-M and rigid-body MolCrystalFlow generators, demonstrating transfer across architectures and representations. Together, our results suggest a simple principle: learn reusable physical alignment into the generator, and reserve inference-time physics for residual constraints that are better corrected than learned.
Related
- On the physics of chromatin molecular complex: from phase-separation to membrane adsorption
- Regenerative Development Goals should shape the post-2030 sustainability agenda
- Extracellular Vesicles in Dermatology: Where Are We and What Lies Ahead?
- Unraveling a Case Where the NAC Approach Fails─MD and QM Modeling of Regioselective cis-Stilbene Oxide Hydrolysis by Engineered Epoxide Hydrolases.
- Crystal structure of the LMAN1-CRD/MCFD2 transport receptor complex provides insight into combined deficiency of factor V and factor VIII.
BibTeX
@misc{tang2026where,
title = {{Where Should Physics Enter a Molecular Crystal Generator?}},
author = {Haocheng Tang and Junmei Wang and Wengong Jin},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.36398},
eprint = {2609.36398},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.36398},
}
RIS
TY - UNPB
TI - Where Should Physics Enter a Molecular Crystal Generator?
AU - Haocheng Tang
AU - Junmei Wang
AU - Wengong Jin
PY - 2026
DA - 2026/09/28
PB - arXiv
DO - 10.48550/arxiv.2609.36398
UR - https://arxiv.org/abs/2609.36398
AB - Generative models make molecular crystal structure prediction fast, but their samples still exhibit geometric and packing violations. Physics can be introduced during training, post-training, or inference, yet these choices are rarely compared with the generator and physical signal held fixed. We introduce CrystAF, an all-atom crystal flow-map generation model, and use it with the UMA interatomic potential to systematically study where physics should enter. Post-training learns physical preferences directly into CrystAF, improving molecular validity and crystal packing while leaving sampling unchanged: physics is paid for once during training rather than repeatedly at deployment. In contrast, UMA relaxation is effective at repairing local clashes but makes generation 6--26× slower, while learning from relaxed targets provides little benefit. These routes are complementary rather than competing. Physics-informed post-training first shifts the generated distribution toward more physically reasonable structures, after which inexpensive inference-time corrections further remove clashes and restore stereochemistry that the generator cannot represent. Importantly, the same post-training strategy also improves the multi-step all-atom Clari-M and rigid-body MolCrystalFlow generators, demonstrating transfer across architectures and representations. Together, our results suggest a simple principle: learn reusable physical alignment into the generator, and reserve inference-time physics for residual constraints that are better corrected than learned.
ER -