- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
The paper studies the problem of inverse design of crystals with target properties, with a technical focus on training-free steering of generative models. The paper argues that existing methods that can steer the inference of a generative model only work on the continuous channels of crystal structures (coordinates and lattices), but do not work on the discrete channels (composition, or atomic elements). The paper proposes FK-ReMask that can steer both continuous and discrete channels during the generation of crystal structures without re-training the model.
Critical strengths
- The paper tackles a relevant research problem in the field of crystal material design with a straightforward method.
Critical weaknesses
- The technical novelty of the proposed method can be limited. On paper, most of the technical designs in the proposed method are derived from existing techniques from either the crystal material design field or other fields (most of which are discussed in the paper's related work, so kudos to the authors for that).
- The proposed method is built on top of MatterGen, and the proposed method is only compared to the vanilla MatterGen, along with a few ablations of it. It could provide more empirical evidence if more conditional crystal structure generation methods could be compared.
Questions for authors
- Could the authors reiterate the technical contribution of their proposed method?
- Could the authors debate the rationale for only comparing with the vanilla MatterGen and its variations in the experiments, or even better, provide experimental comparison with more methods?
3 comments
2. Following the reviewer's suggestion, we extend the comparison in three directions:
-
A second base model with a different conditioning mechanism. ChargeDIFF (Park et al., 2026) jointly generates atomic structures and electron charge densities. Unlike MatterGen's CFG-based conditioning, ChargeDIFF's band-gap-conditional model ("Conditional") directly incorporates the target into the denoiser, without classifier-free guidance.
Applying FK-ReMask (k=8) to unconditional ChargeDIFF achieves a 9.6% SUN & on-target yield, compared with 3.1% for Best-of-N and 2.4% for Conditional, but it is below Conditional + Best-of-N (13.2%).
FK-ReMask can also be applied on top of Conditional. Combined with a novelty reward, it achieves a SUN & on-target yield of 14.8%, outperforming Conditional + Best-of-N (13.2%) and achieving the highest yield among the evaluated methods. The latter combination comes at the cost of reduced validity (64.8%). Adding a validity term restores validity to 84.1%, the highest in the table, while keeping SUN & on-target at 14.7% (Appendix F, Table A6).
| method | valid | unique | novel | stable | SUN | on-target | stable & on-target | SUN & on-target |
|---|---|---|---|---|---|---|---|---|
| ChargeDIFF | 83.9 ± 1.0 | 99.2 ± 0.1 | 64.1 ± 1.3 | 51.8 ± 0.7 | 21.0 ± 0.5 | 0.9 ± 0.4 | 0.4 ± 0.1 | 0.1 ± 0.1 |
| + Best-of-N | 83.7 ± 0.8 | 98.0 ± 0.2 | 69.6 ± 0.9 | 51.7 ± 0.8 | 25.3 ± 0.2 | 10.7 ± 0.7 | 5.7 ± 0.4 | 3.1 ± 0.5 |
| + Best-of-N + novelty | 81.6 ± 0.5 | 98.0 ± 0.1 | 78.6 ± 0.8 | 45.8 ± 0.7 | 27.5 ± 0.4 | 9.4 ± 0.9 | 4.8 ± 0.2 | 2.8 ± 0.5 |
| + Conditional | 82.2 ± 1.0 | 97.6 ± 0.7 | 69.9 ± 0.8 | 52.9 ± 1.8 | 26.3 ± 2.3 | 9.5 ± 1.3 | 6.6 ± 1.3 | 2.4 ± 1.0 |
| + Conditional + Best-of-N | 83.3 ± 0.9 | 91.1 ± 0.3 | 63.5 ± 1.3 | 63.3 ± 0.5 | 28.1 ± 1.3 | 53.4 ± 0.5 | 34.8 ± 0.6 | 13.2 ± 0.7 |
| + Conditional + Best-of-N + novelty | 74.7 ± 1.3 | 95.6 ± 0.3 | 83.6 ± 0.5 | 45.8 ± 0.8 | 30.3 ± 0.5 | 40.2 ± 0.7 | 21.6 ± 0.9 | 10.9 ± 0.6 |
| + FK-ReMask | 79.4 ± 0.1 | 89.1 ± 0.9 | 55.4 ± 1.1 | 65.7 ± 0.5 | 26.0 ± 1.1 | 35.7 ± 0.5 | 24.9 ± 1.1 | 9.6 ± 0.6 |
| + FK-ReMask + novelty | 72.3 ± 1.1 | 92.5 ± 0.7 | 75.1 ± 0.2 | 48.4 ± 0.6 | 26.5 ± 0.4 | 30.1 ± 1.0 | 15.4 ± 0.9 | 6.8 ± 0.5 |
| + Conditional + FK-ReMask | 81.1 ± 0.6 | 77.3 ± 0.5 | 47.7 ± 1.1 | 79.6 ± 1.5 | 29.6 ± 0.1 | 55.3 ± 0.9 | 46.7 ± 1.3 | 12.4 ± 0.2 |
| + Conditional + FK-ReMask + novelty | 64.8 ± 0.3 | 89.2 ± 0.5 | 76.2 ± 0.9 | 55.8 ± 0.8 | 34.0 ± 1.0 | 47.5 ± 0.8 | 30.2 ± 0.5 | 14.8 ± 0.3 |
| + Conditional + FK-ReMask + novelty + validity | 84.1 ± 0.3 | 86.5 ± 1.0 | 72.1 ± 0.8 | 61.0 ± 0.6 | 33.6 ± 0.3 | 48.0 ± 1.2 | 32.8 ± 0.7 | 14.7 ± 0.4 |
-
Additional training-free steering method. We compared FK-ReMask with Universal Guidance (UG). UG steers only atom positions and the lattice, whereas FK-ReMask also steers atom types, enabling control over properties of the entire crystal.
For band gap and bulk modulus tasks, FK-ReMask achieves the highest on-target, stable & on-target, and SUN & on-target yields (single seed, 512 crystals each).
We additionally evaluate a task from the UG paper: generating Li-Co-O crystals with a mean Co-O coordination number of 4. All methods use MatterGen conditioned on Li-Co-O with CFG, which encourages but does not enforce the presence of these elements. Consequently, 80% of vanilla MatterGen’s generated crystals contain no cobalt, making the target inaccessible through atomic-position steering alone. UG does not overcome this limitation (only 19% of its crystals contain cobalt), whereas FK-ReMask increases this fraction to 75% and achieves the highest yields on all three metrics.
| Valid | Unique | Novel | Stable | SUN | On target | Stable & on target | SUN & on target | |
|---|---|---|---|---|---|---|---|---|
| Band gap, 2.4 ± 0.1 eV | ||||||||
| MatterGen | 86.3 ± 1.5 | 99.6 ± 0.3 | 59.2 ± 2.2 | 75.0 ± 1.9 | 37.5 ± 2.1 | 0.4 ± 0.3 | 0.4 ± 0.3 | 0.2 ± 0.2 |
| + Universal Guidance | 86.1 ± 1.5 | 99.8 ± 0.2 | 59.6 ± 2.2 | 76.8 ± 1.9 | 39.6 ± 2.2 | 2.7 ± 0.7 | 2.1 ± 0.6 | 1.2 ± 0.5 |
| + FK-ReMask | 80.9 ± 1.7 | 96.1 ± 0.9 | 68.2 ± 2.1 | 82.6 ± 1.7 | 51.6 ± 2.2 | 64.5 ± 2.1 | 51.6 ± 2.2 | 32.6 ± 2.1 |
| Bulk modulus, 150 ± 15 GPa | ||||||||
| MatterGen | 86.3 ± 1.5 | 99.6 ± 0.3 | 59.2 ± 2.2 | 75.0 ± 1.9 | 37.5 ± 2.1 | 6.4 ± 1.1 | 4.1 ± 0.9 | 1.8 ± 0.6 |
| + Universal Guidance | 86.5 ± 1.5 | 99.4 ± 0.3 | 58.4 ± 2.2 | 77.7 ± 1.8 | 40.0 ± 2.2 | 6.1 ± 1.1 | 5.5 ± 1.0 | 2.0 ± 0.6 |
| + FK-ReMask | 76.0 ± 1.9 | 96.9 ± 0.8 | 52.7 ± 2.2 | 76.8 ± 1.9 | 33.4 ± 2.1 | 96.1 ± 0.9 | 74.0 ± 1.9 | 32.4 ± 2.1 |
| Mean Co–O coordination, Li–Co–O, 4 ± 0.5 | ||||||||
| MatterGen, conditioned on Li–Co–O | 48.8 ± 2.2 | 60.9 ± 2.2 | 78.9 ± 1.8 | 37.7 ± 2.1 | 9.8 ± 1.3 | 5.1 ± 1.0 | 2.9 ± 0.8 | 2.0 ± 0.6 |
| + Universal Guidance | 42.0 ± 2.2 | 65.4 ± 2.1 | 83.4 ± 1.6 | 29.1 ± 2.0 | 10.9 ± 1.4 | 4.5 ± 0.9 | 0.8 ± 0.4 | 0.8 ± 0.4 |
| + FK-ReMask | 74.4 ± 1.9 | 60.7 ± 2.2 | 81.4 ± 1.7 | 38.9 ± 2.2 | 13.3 ± 1.5 | 37.1 ± 2.1 | 9.8 ± 1.3 | 5.7 ± 1.0 |
- FK-ReMask as a complementary tool for conditional models. We applied FK-ReMask on top of MatterGen with classifier-free guidance (CFG), whose property adapter is fine-tuned on labelled data. On bandgap, and bulk modulus, it improves the SUN & on-target rates, in comparison to CFG and CFG + Best-of-N (Appendix G).
| property | method | valid | unique | novel | stable | SUN | on-target | stable & on-target | SUN & on-target |
|---|---|---|---|---|---|---|---|---|---|
| Bandgap (E_g) | CFG | 86.3 ± 0.2 | 99.0 ± 0.2 | 69.0 ± 1.6 | 66.9 ± 1.2 | 37.7 ± 0.6 | 5.0 ± 0.3 | 3.3 ± 0.1 | 1.5 ± 0.3 |
| CFG + Best-of-N | 86.0 ± 1.1 | 97.7 ± 0.2 | 68.9 ± 1.5 | 68.9 ± 1.0 | 39.5 ± 0.4 | 36.5 ± 0.9 | 25.2 ± 0.9 | 13.5 ± 0.4 | |
| CFG + FK-ReMask | 88.2 ± 1.5 | 90.7 ± 0.2 | 62.1 ± 1.0 | 86.1 ± 0.9* | 47.3 ± 0.5* | 63.3 ± 0.4* | 55.7 ± 0.8* | 29.9 ± 0.3* | |
| Bulk Modulus (K) | CFG | 81.2 ± 1.4 | 99.5 ± 0.1 | 57.1 ± 0.8 | 66.1 ± 1.3 | 27.9 ± 1.1 | 53.3 ± 0.3 | 36.8 ± 0.5 | 15.8 ± 0.2 |
| CFG + Best-of-N | 78.7 ± 0.6 | 98.8 ± 0.2 | 55.2 ± 0.1 | 65.9 ± 1.3 | 27.3 ± 0.7 | 99.7 ± 0.1 | 65.7 ± 1.4 | 27.2 ± 0.8 | |
| CFG + FK-ReMask | 80.8 ± 0.5 | 95.3 ± 0.5 | 52.7 ± 1.5 | 76.4 ± 0.7* | 33.7 ± 1.4 | 99.7 ± 0.2 | 76.2 ± 0.6* | 33.5 ± 1.5* |
We thank the reviewer for the feedback and for appreciating our effort to position the proposed method in the context of existing work.
1. Our technical contribution: