- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies why diffusion models can generate structures that match the observed data but still contain physical defects. The authors make four changes to the GLASS sampling process without retraining the model. Experiments on amorphous silicon show lower excess entropy and improvements in several structural metrics.
Critical strengths
1.The entropy analysis provides a useful way to assess individual structures. Its relationship with defects and energy helps connect the evaluation metric to physical quality.
2.The method improves several structural metrics without retraining the score network.
3.The authors provide component ablations and evaluate energy using a different potential from the one used for relaxation. These are useful checks.
Critical weaknesses
1.The method is only evaluated with GLASS, with no direct comparison to other generation or refinement methods. It is unclear whether existing refinement methods could achieve similar improvements or whether the proposed changes would help other diffusion models.
2.In the defect-free experiment, the baseline is unconditional, while the proposed method uses additional guidance. Appendix A.7 also changes the observable resolution and refinement budget from the standard recipe. These differences make it difficult to identify what drives the improvements in Figure 4(d–f).
3.Some entropy comparisons use relaxed generated samples and unrelaxed references. Appendix A.2 also shows that displacement noise explains only part of the entropy gap. These results do not establish residual temperature as the main cause of the bias.
Questions for authors
1.How does the method compare with GLASS using its existing annealing and refinement options under the same information and compute budget? Have you tested another generation or refinement method on a comparable task?
2.Can you repeat the defect-free comparison with the same conditioning information, observable resolution, and compute budget for both methods?
3.Which data are used to construct the low-frequency S(k) target? How does the method perform when only the truncated g2(r) is available?
4.How do the entropy, energy, and structural comparisons change when the baseline outputs, proposed outputs, and reference structures are all evaluated under the same relaxation conditions?
7 comments
Reply to Reviewer iR9D: Question 1
1.How does the method compare with GLASS using its existing annealing and refinement options under the same information and compute budget? Have you tested another generation or refinement method on a comparable task?
Our choice of GLASS was motivated by our internal analysis of the landscape of generative models for amorphous structures, where GLASS was the strongest performer across many models and metrics of success. We had tested multiple models in the literature, including amorphous-focused DM2 (Yang and Schwalbe-Koda, 2026), AMDEN (Finkler et al., 2026), GLASS (Guo and Schwalbe-Koda, 2026), as well as adapted flagship crystal-focused models such as CDVAE (Xie et al., 2022) and MatterGen (Zeni et al., 2025) to generating amorphous materials based on their structures. Across this benchmark, which comprised over 20 quality metrics and 5 systems (a-C, a-Si, a-SiO2, CuZr, NiAlZr) under different conditions (processing, density, composition), GLASS was the strongest performer. GLASS ranks first in every system, a-Si and CuZr included, at 1.3x the MD noise floor. In contrast, results from baseline reverse Monte Carlo and hybrid RMC methods are at least 8x and 13x worse than the MD simulations.
Below, we reproduce two tables of that study for the Reviewer’s information.
Model rankings across metrics for a-Si
| Metric | RMC | HRMC | CRN | GLASS | DM2 | CDVAE | MatterGen | AMDEN |
|---|---|---|---|---|---|---|---|---|
| g(r) | 24 ±0.2 | 33 ±0.2 | 13 ±1.0 | 0.5 ±0.1 | 3.7 ±0.4 | 1.1 ±0.2 | 10 ±2.5 | 12 ±1.1 |
| ADF | 21 ±0.3 | 26 ±0.2 | 4.5 ±0.1 | 0.8 ±0.2 | 4.0 ±0.2 | 0.8 ±0.1 | 7.7 ±0.6 | 18 ±0.4 |
| CN dist | 7.2 ±0.1 | 8.6 ±0.2 | 3.7 ±0.3 | 0.8 ±0.1 | 1.0 ±0.1 | 1.0 ±0.1 | 4.2 ±0.4 | 4.8 ±0.3 |
| CN outliers | 46 ±2.6 | 65 ±3.5 | 0.0 ±0.0 | 1.3 ±0.4 | 1.3 ±0.3 | 0.7 ±0.3 | 8.0 ±2.3 | 17 ±1.8 |
| mean CN | 2.8 ±0.1 | 10 ±0.4 | 3.4 ±0.1 | 0.4 ±0.1 | 0.3 ±0.3 | 0.8 ±0.1 | 7.9 ±0.9 | 5.9 ±0.6 |
| rings | 5.2 ±0.5 | 1.8 ±0.3 | 2.8 ±0.3 | 0.7 ±0.2 | 0.9 ±0.3 | 0.5 ±0.1 | 7.5 ±1.2 | 0.8 ±0.6 |
| ring dist | 6.8 ±0.1 | 8.0 ±0.2 | 1.7 ±0.1 | 0.8 ±0.1 | 1.6 ±0.1 | 1.0 ±0.1 | 2.3 ±0.1 | 5.4 ±0.2 |
| ring outliers | 23 ±4.2 | 18 ±4.0 | 0.0 ±0.4 | 0.0 ±0.1 | 0.0 ±0.1 | 0.0 ±0.0 | 0.0 ±0.4 | 14 ±3.2 |
| Voronoi | 1.9 ±0.0 | 2.2 ±0.0 | 1.1 ±0.0 | 1.0 ±0.0 | 1.0 ±0.0 | 0.9 ±0.1 | 1.0 ±0.0 | 1.6 ±0.0 |
| XRD | 1.9 ±0.3 | 2.5 ±0.1 | 3.5 ±0.4 | 1.6 ±0.2 | 1.8 ±0.4 | 1.3 ±0.7 | 9.7 ±2.2 | 2.2 ±0.3 |
| ND | 1.9 ±0.2 | 2.2 ±0.1 | 2.8 ±0.3 | 1.3 ±0.2 | 1.4 ±0.3 | 1.1 ±0.6 | 8.7 ±2.0 | 1.8 ±0.1 |
| bulk K | 39 ±2.4 | 40 ±206 | 13 ±1.2 | 3.5 ±0.6 | 13 ±1.7 | 5.7 ±1.0 | 5.7 ±4.7 | 29 ±2.2 |
| shear G | 15 ±1.5 | 18 ±1.3 | 5.7 ±0.6 | 3.6 ±0.9 | 9.3 ±1.4 | 5.7 ±0.8 | 2.1 ±2.4 | 14 ±0.9 |
| Young E | 17 ±1.6 | 19 ±1.3 | 6.1 ±0.6 | 3.6 ±1.0 | 9.9 ±1.4 | 5.8 ±0.8 | 2.9 ±2.1 | 15 ±1.0 |
| relax dist | 26 ±0.7 | 31 ±1.1 | 0.7 ±0.3 | 0.6 ±0.2 | 6.4 ±0.8 | 0.9 ±0.3 | 1.3 ±0.4 | 23 ±0.5 |
| relaxed n [converged/total] | 10/10 | 10/10 | 10/10 | 10/10 | 10/10 | 8/10 | 5/5 | 10/10 |
| yield σ | 1.4 ±0.1 | 1.4 ±0.1 | 1.3 ±0.5 | 0.7 ±0.2 | 1.0 ±0.1 | 0.5 ±0.6 | 1.0 ±0.3 | 1.5 ±0.3 |
| UTS | 6.9 ±0.4 | 6.8 ±0.4 | 1.6 ±0.4 | 1.5 ±0.6 | 2.8 ±0.2 | 0.8 ±0.4 | 4.0 ±1.1 | 6.4 ±0.3 |
| Energy/atom | 136 ±1.7 | 172 ±2.0 | 6.1 ±1.3 | 3.7 ±0.4 | 16 ±0.8 | 2.1 ±0.3 | 6.2 ±0.4 | 80 ±2.1 |
| Forces | 64 ±0.6 | 68 ±0.9 | 19 ±1.3 | 1.2 ±0.3 | 5.5 ±0.4 | 4.4 ±0.4 | 8.6 ±0.5 | 30 ±1.7 |
| QUESTS | 7.3 | 7.9 | 1.7 | 1.1 | 1.0 | 1.1 | 3.1 | 4.8 |
| QUESTS div | 14 | 15 | 1.0 | 1.1 | 0.1 | 0.9 | 5.7 | 12 |
| foreign n [count/total] | 0/10 | 0/10 | 0/10 | 0/10 | 0/10 | 2/10 | 0/5 | 0/10 |
| Overall (×MD) | 14 | 15 | 2.8 | 1.1 | 1.6 | 1.0 | 5.7 | 12 |
| Mean rank | 6.7 | 7.6 | 4.1 | 1.8 | 3.1 | 2.4 | 4.5 | 5.7 |
Overall model rankings across systems
| System | RMC | HRMC | CRN | GLASS | DM2 | CDVAE | MatterGen | AMDEN |
|---|---|---|---|---|---|---|---|---|
| a-C | 5.4 | 6.2 | — | 1.5 | 2.8 | 2.6 | 3.1 | 5.0 |
| a-Si | 6.4 | 7.1 | 5.5 | 1.8 | 3.0 | 2.6 | 4.1 | 5.4 |
| SiO2 | 6.6 | 7.1 | 4.1 | 2.1 | 3.4 | 2.6 | 4.6 | 5.5 |
| CuZr | 4.9 | — | — | 1.9 | 2.4 | 3.4 | 3.8 | 4.6 |
| ZrNiAl | 5.4 | 5.9 | — | 2.6 | 3.0 | 3.2 | 4.1 | 3.9 |
| mean rank | 5.7 ±0.04 | 6.6 ±0.04 | 4.8 ±0.06 | 2.0 ±0.07 | 2.9 ±0.06 | 2.9 ±0.11 | 4.0 ±0.08 | 4.9 ±0.06 |
| mean ×MD | 8.2 | 13 | 5.7 | 1.3 | 2.2 | 1.7 | 3.7 | 8.2 |
Whereas reproducing the entire benchmark with our method is not the main focus of this manuscript, our focus on GLASS essentially advances the state-of-the-art in quality of generated amorphous structures. We believe that the same approach would be applicable to other models, our selection of GLASS was meant to start with the best-performing model, and diagnose the contributions of our sampling strategies plus corrections.
Reply to Reviewer iR9D: Question 2
2.Can you repeat the defect-free comparison with the same conditioning information, observable resolution, and compute budget for both methods?
We thank the reviewer for pointing this out. To clarify this point, we have updated Figure 4 to have all results based on a conditional generation strategy with identical settings of guidance information, resolution, and budget - except, of course, for those that comprise our annealing strategy. In Figs. 4a-c, both runs are now conditioned on the same : "GLASS, no recipe" is GLASS as-published with the corresponding guidance, and the “GLASS + annealing” includes our strategy. The results are nearly identical to before, and the conclusions remain true: our annealing strategy continues to provide improvements even when compared to the guided diffusion.
In the original Figs. 4d-f, the experiments were initially performed with no guidance term because the prior had been trained on the target structure alone, which aimed to be replicated with the generative model. Therefore, the unconditionally generated configurations already had near identical to the target one. Nevertheless, we agree with the reviewer that a reasonable comparison requires measuring both against the same result. Quantitatively, adding the guidance alone changes the entropy only by 0.04 nats, thus within the original error bar of generated configurations. This further demonstrates that the difference between the two arms comes from the modifications and not from the guidance. We clarified this in the caption of Fig. 4.
Regarding the budget at inference time, GLASS (no recipe) and GLASS + annealing, by design, have different computational budgets. Even though both tests have been performed with the same number of denoising steps, model sizes, and all other fixed hyperparameters, the main goal of this work is to improve the quality of the generated amorphous structures at the expense of computational budget. Within amorphous materials, the inference cost of a generative model is negligible compared to the cost of producing reference structures (which can be way over 4 orders of magnitude more expensive than a single inference). Other complicating factors, such as the choice of MLIP used for the projected relaxation, could also change the wall time for inference, which would create issues when using “constant budget” as constraint to evaluate the models. Given that most ML research is performed to maximize metrics even at the expense of computational cost, we believe this approach is reasonable.
Nevertheless, we compared the controlled annealing from our strategy to simply increasing (or restarting) the diffusion process without the other terms. Whereas the method discussed in Section 5 of our manuscript led to clear improvements in structure quality, simple restarts led to structural quality metrics statistically indistinguishable from the original recipe.
To further demonstrate that, using GLASS with the original step count (1074 steps over four passes) produces structures with H = 3.73 nats, while the annealing with the new settings leads to H = 3.38 nats with the new budget. The defect-free target is a harder object than the MD ensemble, a single structure whose formation by melt-quench its authors describe as a rare event, so what it needed to approach its reference was not a different method but a larger budget. The annealing therefore trades compute for entropy, within a limit set by the complexity of the structures sought. This discussion was added to the Appendix A.9.
Reply to Reviewer iR9D: Overview and Critical Weaknesses
We thank the reviewer for the comments. Below, we address each of the questions and the weakness highlighted. The manuscript PDF was updated to include all the new information and analysis from this discussion. Modifications to the manuscript are shown in red color.
Reviewer iR9D highlighted the following critical weaknesses:
For each of these, the reviewer asked follow-up questions that address each, so each perceived weakness has been mitigated by additional data, experiments, and the discussion below.