Challenges in Non-Polymeric Crystal Structure Prediction: Why a Geometric, Permutation-Invariant Loss is Needed
Organizations: Univ. Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK, 38000 Grenoble, France. · Univ. Grenoble Alpes, CNRS, Grenoble INP, LJK, 38000 Grenoble, France.
Abstract
Crystalline structure prediction is an essential prerequisite for designing materials with targeted properties. Yet, it is still an open challenge in materials design and drug discovery. Despite recent advances in computational materials science, accurately predicting three-dimensional non-polymeric crystal structures remains elusive. In this work, we focus on the molecular assembly problem, where a set~ of identical rigid molecules is packed to form a crystalline structure. Such a simplified formulation provides a useful approximation to the actual problem. However, while recent state-of-the-art methods have increasingly adopted sophisticated techniques, the underlying learning objective remains ill-posed. We propose a better formulation that introduces a loss function capturing key geometric molecular properties while ensuring permutation invariance over . Remarkably, we demonstrate that within this framework, a simple regression model already outperforms prior approaches, including flow matching techniques, on the COD-Cluster17 benchmark, a curated non-polymeric subset of the Crystallography Open Database (COD). Our code is available at https://github.com/inria-thoth/SinkFast.
Figures & tables
| Pretrained Weights | Flow Matching | Packing Matching in Å | ||||
| Dataset: COD-Cluster17-5K | ||||||
| PackMol ( Martínez et al., 2009 ) | - | - | ||||
| - | - | - | ||||
| - | - | - | ||||
| ✓ | - | - | - | |||
| Test Loss in Å | Packing matching in Å | ||||||
| Loss | Flow Matching | ||||||
| Dataset: COD-Cluster17-5K | |||||||
| ✓ | |||||||
| ✓ | |||||||
| ✓ | |||||||
| Method | Training time (s) per epoch | Test time (s) per batch |
| Cond-CDVAE - pretrained | - | 5760 |
| AssembleFlow | 2678.5 | 0.89 |
| SinkFast - without permutation-invariant loss | 53.6 | 0.02 |
| SinkFast - with permutation-invariant loss | 64.1 | 0.02 |
Appendix figures & tables26 assets
Supplementary material from the paper’s appendix.
Appendix
| def AtomModel( ) |
| 1: t=\text{Linear}(\text{SiLU}(\text{Linear}(\text{time_embed}(t)))) [c] |
| 2: [ ] |
| 3: [ ] |
| 4: [ ] |
| 5: Atom to Molecules edges |
| 6: for all |
| Model part | Function | Parameters |
| Training | Epochs | {500} |
| Batch size | {8} | |
| Loss | {LM: {alpha: 10}} | |
| {RMSD: } | ||
| {Geometric: } | ||
| Assignment | {None: } |
| Package | Version | License |
| COD-Cluster17 | git commit | MIT |
| AssembleFlow Model | git commit | MIT |
| POT | 0.9.5 | MIT |
| RMSD | - | CeCILL |
| Test Loss in Å | Packing matching in Å | ||||||
| Loss | Diff. assign. | ||||||
| Dataset: COD-Cluster17-5K | |||||||
| ✓ | |||||||
| Test Loss in Å | Packing matching in Å | ||||||
| Loss | Flow Matching | ||||||
| Dataset: COD-Cluster17-5K | |||||||
| ✓ | |||||||
| ✓ | ||||
| ✓ | ||||
| ✓ | ✓ |
| ✓ | ||||
| ✓ | ||||
| ✓ | ✓ |
| Test Loss in Å | Packing matching in Å | ||||||
| Loss | Method | ||||||
| Dataset: COD-Cluster17-5K | |||||||
| AssembleFlow | |||||||
| AssembleFlow | |||||||
| AssembleFlow | |||||||
| SinkFast | |||||||
| Test Loss in Å | Packing matching in Å | ||||||
| Loss | Assignment type | ||||||
| Dataset: COD-Cluster17-5K | |||||||
| Exact | |||||||
| Exact | |||||||
| Exact | |||||||
| Diff. | |||||||
| Test Loss in Å | ||||
| Model | Loss | |||
| Dataset: COD-Cluster17-5K | ||||
| Noise (Baseline) | ||||
| AssembleFlow | ||||
| def AtomModel( |
| ) |
| 1 : t=\text{MLP}(\text{time_embed}(t)) [c] |
| 2 : [ ] |
| 3 : [ ] |
| 4 : [ ] |
| 5 : Atom to Molecules edges |
| def |
| ) |
| 2: [Edges, 1] |
| 3: with the outer product [ ] |
| 4: [ ] |
| 3: get dominant eigen vector [ ] |
| 5: |
| Flow Matching | Equivariant Backbone | |||
| 8.87 | 5.48 | 6.53 | ||
| ✓ | 9.22 | 5.55 | 6.66 | |
| ✓ | 10.34 | 3.99 | 5.88 | |
| ✓ | ✓ | 10.26 | 4.79 | 6.27 |
| Sinkhorn Regularization Parameter | |||
| 100 | |||
| 50 | |||
| 10 | |||
| 5 | |||
| 2 | |||
| 1 |
| Method | ||
| CDVAE | ||
| DiffCSP | ||
| AssembleFlow | ||
| SinkFast |
| Method | ||
| CDVAE | ||
| DiffCSP | ||
| AssembleFlow | ||
| SinkFast |
| Method | ||
| CDVAE | ||
| DiffCSP | ||
| AssembleFlow | ||
| SinkFast |
| Method | ||
| CDVAE | ||
| DiffCSP | ||
| AssembleFlow | ||
| SinkFast |
| Method | ||
| CDVAE | ||
| DiffCSP | ||
| AssembleFlow | ||
| SinkFast |
| Method | ||
| CDVAE | ||
| DiffCSP | ||
| AssembleFlow | ||
| SinkFast |
| Method | RDKit | ||
| AssembleFlow | |||
| AssembleFlow | ✓ | ||
| SinkFast | |||
| SinkFast | ✓ |
| Method | RDKit | ||
| AssembleFlow | |||
| AssembleFlow | ✓ | ||
| SinkFast | |||
| SinkFast | ✓ |
| Method | RDKit | ||
| AssembleFlow | |||
| AssembleFlow | ✓ | ||
| SinkFast | |||
| SinkFast | ✓ |
Crystal Structure Prediction
Crystal structure prediction seeks stable atomic arrangements from composition, molecular identity, or experimental measurements.
See all 38 papers