acceptodds
Under review as a conference paper at ICLR 2027

Breaking Observable Degeneracy in Diffusion Models of Amorphous Matter

14 views

Abstract

Diversity in conditional samples from a generative model can reflect either creativity or hallucination. In the physical sciences, this ambiguity is critical when inverting an experimental measurement, since choosing among the solutions requires an intractable ground truth that the generative model is meant to replace. Using disordered materials as a challenging case study, we show that a score-based denoiser systematically produces samples of high configurational entropy that the conditional input cannot distinguish from the reference. This amounts to a residual excess temperature, raising the concentration of outliers in the system and compromising the physical behavior of the samples. To break this degeneracy, we implement four modifications to the sampling routine that suppress this residual noise: a gate interpolating between the reverse SDE and the probability-flow ODE; guidance on the low-frequency bands that the experimental observable leaves undetermined; a structural relaxation projected onto the null space of the observable's Jacobian; and a ladder of partial restarts. Together, these measures halve the excess entropy on benchmark systems and allow reaching amorphous configurations that quench kinetics have kept out of reach of direct simulation for decades.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
generative modeling, probabilistic methods, sampling, physical sciences, materials science

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_iR9DOpenReview ↗
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
AuthorsOpenReview ↗

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:

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.

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.

AuthorsOpenReview ↗

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

MetricRMCHRMCCRNGLASSDM2CDVAEMatterGenAMDEN
g(r)24 ±0.233 ±0.213 ±1.00.5 ±0.13.7 ±0.41.1 ±0.210 ±2.512 ±1.1
ADF21 ±0.326 ±0.24.5 ±0.10.8 ±0.24.0 ±0.20.8 ±0.17.7 ±0.618 ±0.4
CN dist7.2 ±0.18.6 ±0.23.7 ±0.30.8 ±0.11.0 ±0.11.0 ±0.14.2 ±0.44.8 ±0.3
CN outliers46 ±2.665 ±3.50.0 ±0.01.3 ±0.41.3 ±0.30.7 ±0.38.0 ±2.317 ±1.8
mean CN2.8 ±0.110 ±0.43.4 ±0.10.4 ±0.10.3 ±0.30.8 ±0.17.9 ±0.95.9 ±0.6
rings5.2 ±0.51.8 ±0.32.8 ±0.30.7 ±0.20.9 ±0.30.5 ±0.17.5 ±1.20.8 ±0.6
ring dist6.8 ±0.18.0 ±0.21.7 ±0.10.8 ±0.11.6 ±0.11.0 ±0.12.3 ±0.15.4 ±0.2
ring outliers23 ±4.218 ±4.00.0 ±0.40.0 ±0.10.0 ±0.10.0 ±0.00.0 ±0.414 ±3.2
Voronoi1.9 ±0.02.2 ±0.01.1 ±0.01.0 ±0.01.0 ±0.00.9 ±0.11.0 ±0.01.6 ±0.0
XRD1.9 ±0.32.5 ±0.13.5 ±0.41.6 ±0.21.8 ±0.41.3 ±0.79.7 ±2.22.2 ±0.3
ND1.9 ±0.22.2 ±0.12.8 ±0.31.3 ±0.21.4 ±0.31.1 ±0.68.7 ±2.01.8 ±0.1
bulk K39 ±2.440 ±20613 ±1.23.5 ±0.613 ±1.75.7 ±1.05.7 ±4.729 ±2.2
shear G15 ±1.518 ±1.35.7 ±0.63.6 ±0.99.3 ±1.45.7 ±0.82.1 ±2.414 ±0.9
Young E17 ±1.619 ±1.36.1 ±0.63.6 ±1.09.9 ±1.45.8 ±0.82.9 ±2.115 ±1.0
relax dist26 ±0.731 ±1.10.7 ±0.30.6 ±0.26.4 ±0.80.9 ±0.31.3 ±0.423 ±0.5
relaxed n [converged/total]10/1010/1010/1010/1010/108/105/510/10
yield σ1.4 ±0.11.4 ±0.11.3 ±0.50.7 ±0.21.0 ±0.10.5 ±0.61.0 ±0.31.5 ±0.3
UTS6.9 ±0.46.8 ±0.41.6 ±0.41.5 ±0.62.8 ±0.20.8 ±0.44.0 ±1.16.4 ±0.3
Energy/atom136 ±1.7172 ±2.06.1 ±1.33.7 ±0.416 ±0.82.1 ±0.36.2 ±0.480 ±2.1
Forces64 ±0.668 ±0.919 ±1.31.2 ±0.35.5 ±0.44.4 ±0.48.6 ±0.530 ±1.7
QUESTS7.37.91.71.11.01.13.14.8
QUESTS div14151.01.10.10.95.712
foreign n [count/total]0/100/100/100/100/102/100/50/10
Overall (×MD)14152.81.11.61.05.712
Mean rank6.77.64.11.83.12.44.55.7

Overall model rankings across systems

SystemRMCHRMCCRNGLASSDM2CDVAEMatterGenAMDEN
a-C5.46.2—1.52.82.63.15.0
a-Si6.47.15.51.83.02.64.15.4
SiO26.67.14.12.13.42.64.65.5
CuZr4.9——1.92.43.43.84.6
ZrNiAl5.45.9—2.63.03.24.13.9
mean rank5.7 ±0.046.6 ±0.044.8 ±0.062.0 ±0.072.9 ±0.062.9 ±0.114.0 ±0.084.9 ±0.06
mean ×MD8.2135.71.32.21.73.78.2

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.

AuthorsOpenReview ↗

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 g(r)g(r): "GLASS, no recipe" is GLASS as-published with the corresponding g(r)g(r) 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 g(r)g(r) 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 g(r)g(r) 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.

Reviewer_6U5nOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work tackles an important problem in using diffusion models to generate samples of atomic configurations in disordered materials like amorphous silicon (a-Si). Though the model is trained on realistic molecular dynamics (MD) simulations, and the sampling is guided to replicate a particular physical observable (for example the pair distribution function g2(r)g_2(r)), the generated distribution can be markedly different in other observables relative to the training set. The underlying issue is the degeneracy of the configuration space: many atomic configurations, even physically unlikely ones, can exhibit nearly the same g2(r)g_2(r) as realistic configurations. The authors use a-Si datasets to first characterize the problem (showing how configurations with the same g2(r)g_2(r) can be distinguished via a recently introduced entropy measure) and then develop a modified sampling technique that substantially improves sample quality.

Critical strengths

The paper is very nicely written and organized, providing a succinct introduction to the relevant topics (i.e. Henderson's theorem) in a form accessible to a broad audience. The modified sampling approach was well motivated, with the contributions of different added steps clearly delineated (and their necessity spelled out). For the two examples shown (training in one case on an a-Si dataset with multiple configurations, and in the other on one defect-free a-Si structure) the results are impressive: basically halving the gap in observables like entropy between the training data and the generated structures. This is not yet a complete solution to the problem, but a big step forward.

Critical weaknesses

There were no significant weaknesses in the paper. Ideally it would have been nice to see examples from other materials rather than a-Si, but this may be better suited for a follow-up.

Questions for authors

There are a number of points for clarification:

  1. The form of the kernel-density estimate of entropy in Eq. 5 is not immediately obvious without looking into the Schwalbe-Koda 2025 paper. It might be helpful to have a brief motivation of this form (in the spirit of Section A.2 in the SI of that paper) to make the current work more self-contained. This motivation could be included as an appendix if there is not enough space in the main text.

  2. Providing some more details of the temperature sensitivity / insensitivity shown in Fig. 3 might be useful. For example why is Dg2D_{g_2} initially flat with increasing σ\sigma? Is this a rigorous property of Dg2D_{g_2}? In other words, can one mathematically prove that the derivative dDg2/dσ=0dD_{g_2}/d\sigma = 0 at σ=0\sigma =0? Or is it more of a phenomenological observation? And is isotropic Gaussian white noise the only form of noise for which this is true?

  3. The authors use both a larger a-SI MD dataset and the smaller defect-free one in the last section. Since terms like "low defect concentration" are used in the text, it would be helpful to explain more fully how these datasets are distinct (how is the defect concentration quantified, and what is the defect concentration in the first case vs. the second).

6 comments
AuthorsOpenReview ↗

Reply to Reviewer 6U5n: Overview and Critical Weakness

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.

There were no significant weaknesses in the paper. Ideally it would have been nice to see examples from other materials rather than a-Si, but this may be better suited for a follow-up.

We thank the reviewer for this assessment. As mentioned to Reviewer a7gs, we extended our work beyond a single example (within the time constraints of this discussion period) by proving its use on CuZr, a metallic glass with two species and completely different structure compared to amorphous silicon. We followed the paper's protocol regarding training the prior and performing an ablation with and without the recipe against the twelve-metric benchmark. Our annealing recipe lowers the entropy H from 4.46 to 4.29 nats (approaching the reference MD results of 4.17 nats) and drastically improves the energy E from -5.73 to -6.26 eV/atom (almost comparable to MD’s -6.31 eV/atom). The resulting g(r)g(r) mismatch is superior to MD's own variance, and our results improve 11 of 12 benchmark metrics against the original generative model. This shows that our framework and interventions in spectral terms, projected relaxation, and annealing ladder transfer to other materials systems. These results have been added to the Appendix A.7 of the updated manuscript, and briefly mentioned in the main text.

AuthorsOpenReview ↗

Reply to Reviewer 6U5n: Question 1

  1. The form of the kernel-density estimate of entropy in Eq. 5 is not immediately obvious without looking into the Schwalbe-Koda 2025 paper. It might be helpful to have a brief motivation of this form (in the spirit of Section A.2 in the SI of that paper) to make the current work more self-contained. This motivation could be included as an appendix if there is not enough space in the main text.

Thank you for the suggestion. We added it to the Section 4 (header “Breaking the degeneracy with information entropy.”), following SI A.2 of Schwalbe-Koda et al. (2025). The description was also briefly included in the main text of the manuscript. Each atom is described by a vector of cutoff-weighted inverse distances to its neighbours and between them, so a structure is a sample of N such vectors. Eq. 5 is the Shannon entropy of their distribution with the density estimated by a Gaussian kernel of bandwidth h over the same N points; with a kernel bounded in [0, 1], H lies between 0 (all environments identical) and log N (all distinct)

AuthorsOpenReview ↗

Reply to Reviewer 6U5n: Question 2

Providing some more details of the temperature sensitivity / insensitivity shown in Fig. 3 might be useful. For example why is Dg2D_{g2} initially flat with increasing σ\sigma? Is this a rigorous property of Dg2D_{g2}? In other words, can one mathematically prove that the derivative dDg2/dσ=0dD_{g2}/d\sigma = 0 at σ=0\sigma = 0? Or is it more of a phenomenological observation? And is isotropic Gaussian white noise the only form of noise for which this is true?

The temperature insensitivity can be approximated directly, and Gaussian noise is not special. For any displacement uu with zero mean and variance σ2\sigma^2, the noise average of any smooth function of the coordinates is f(x)+(σ2/2)∇2f+O(σ4)f(x) + (\sigma^2 / 2) \nabla^2 f + O(\sigma^4): the term linear in uu averages to zero whatever its distribution. Applied to g2g_2, the change is of order σ2\sigma^2, and since Dg2D_{g2} depends on σ\sigma only through σ2\sigma^2, dDg2/dσ=0dD_{g2}/d\sigma = 0 at σ=0\sigma = 0 for every zero-mean noise, Gaussian or not. Dg2D_{g2} stays flat for longer than H because Dg2D_{g2} is quadratic in that change and rises as σ4\sigma^4 on average over the references, while the kernel density behind H broadens with the jitter and H rises as σ2\sigma^2. That difference of orders is compatible with the window of noises in Fig. 3. We added this derivation and explanation to Appendix A.3 of the manuscript. Regarding Gaussian noise, we also tried other temperature schedules and more structured displacement fields, but all results are nearly indistinguishable for reasons that remain unclear. Nevertheless, isotropic Gaussian noise is a reasonable model of an effective temperature, and is identical to the Einstein model for heat capacity in solids. Therefore, this is the model that motivated our strategy for annealing. A better model with correlated displacements dependent on each atomic environment, for instance, could create a better annealing schedule in future works.

Reviewer_a7gsOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper identifies a practical degeneracy that exists while using diffusion based methods to solve the inverse problem of predicting the atomic structure from the g2g_2 measurements. The authors suggest a variety of strategies to break this degeneracy in practice, including modifying the noising schedule, adding long-wavelength structural constraints etc. The authors test the effect of these using a local entropy function to determine if excess entropy is removed while keeping the g2g_2 close to the target.

Critical strengths

This is a very clearly written paper. The core issues are well presented and the various interventions that the authors use to fix it are well motivated in the text.

Critical weaknesses

While the paper starts off by identifying a very general problem of practical degeneracy of inverse problems that are solved using diffusion based methods, it solves it in a very narrow application. So it is unclear, if the techniques here are general enough to correctly bias these generative models to physically relevant solutions in a more general setting. If it only works for this particular application of predicting atomic structures from g2g_2, then the paper may be better suited for a more specialized venue. A different application would have substantially strengthened the paper. Right now the paper is a very detailed, single case study.

Questions for authors
  1. Does lower entropy always mean a more physically relevant structure? I would imagine, this would require a more thorough characterization of the generated structure, which would include the entropy as one of the components.

  2. Related to the above question: when several physically plausible structures reproduce the same measurement, should the goal be to represent that uncertainty or to select a low-defect structure? What additional physical assumptions justify preferring lower-entropy solutions? I can imagine, in other physical situations, may be selecting a solution that satisfies some known symmetries of the problem would be the right approach.

  3. Adding random displacements reproduces some signatures of the generated structures, but does that establish that residual thermal noise causes the observed defects? How can this explanation be distinguished from an inaccurate learned prior?

6 comments
AuthorsOpenReview ↗

Reply to Reviewer a7gs: Overview and Critical Weakness

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.

[this paper] solves it in a very narrow application. So it is unclear, if the techniques here are general enough to correctly bias these generative models to physically relevant solutions in a more general setting. [...] A different application would have substantially strengthened the paper.

We thank the reviewer for this comment. Whereas the application space is seemingly limited from a machine learning perspective, the problem of solving structures from g2(r)g_2(r), especially in amorphous matter, is one of the most persistent and impactful issues in the intersection of sampling and materials science. Currently, the vast majority of materials characterization problems that use g2(r)g_2(r) rely on manual interpretation of data to solve this problem - and these challenges are ubiquitous. The case is even worse in amorphous matter, where little intuition is available for experts and is exacerbated by the vast structural space available for sampling. In these systems, the lack of long-range drastically worsen the structure sampling problem compared to inorganic crystals. Therefore, our work tackles some of the worst-case scenarios of this problem with a combination of theory and empirical results that are expected to generalize well across the space of materials. As an additional point, the discussion in the text highlights a sampling problem in a-Si that lingered for decades, exemplifying a case where scientific impact in this domain specific field is limited not by the number of samples, but by their quality as well.

To extend this beyond a single example (within the time constraints of this discussion period) we demonstrate our method on CuZr, a metallic glass with two species and completely different structure compared to the initial amorphous silicon. We followed the paper's protocol regarding training the prior and performing an ablation with and without the recipe against the twelve-metric benchmark. Our annealing recipe lowers the entropy H from 4.46 to 4.29 nats (approaching the reference MD results of 4.17 nats) and drastically improves the energy E from -5.73 to -6.26 eV/atom (almost comparable to MD’s -6.31 eV/atom). The resulting g(r)g(r) mismatch is superior to MD's own variance, and our results improve 11 of 12 benchmark metrics against the original generative model. Therefore, our framework and interventions in spectral term, projected relaxation, and annealing ladder transfer to other materials systems. These results have been added to the Appendix A.7 of the manuscript.

As an additional point on the generality of the method, our annealing scheme does not change the trained models. Rather, we provide a theoretical viewpoint that relates structural quality and observable degeneracy, and propose strategies that improve the generated samples. The methods may prove useful to other models in the sampling community across different systems in follow up works.

AuthorsOpenReview ↗

Reply to Reviewer a7gs: Question 1

  1. Does lower entropy always mean a more physically relevant structure? I would imagine, this would require a more thorough characterization of the generated structure, which would include the entropy as one of the components.

At a fixed observable, we postulate that physically meaningful structures often carry less information. Loosely borrowing terms from statistical thermodynamics, there are many more microstates at higher energy than lower energy, and sampling is often useful to obtain structures at lower temperature. Schwalbe-Koda et al. (2025) developed this parsimony reading for the entropy estimator and found it empirically consistent with this viewpoint. The opposite tendency is the known failure of reverse Monte Carlo, the traditional structure inversion method which returns the most disordered structure consistent with the data. In our manuscript, we formulate theoretical justifications to this: Fig. 2 shows that in this parameter, the sampler's iso-g2g_2 samples and the MD references occupy different distributions. Entropy condenses many of the system-specific observables (e.g., ring size distribution) into one actionable quantity that is simpler to track and modulate compared to geometrical ones. Thus, while we have not proven that the entropy minimization is the only way to obtain physically meaningful structures at constant g2(r)g_2(r), we have showed theoretical arguments and empirical evidence that this is a reasonable strategy for now.

AuthorsOpenReview ↗

Reply to Reviewer a7gs: Question 2

  1. Related to the above question: when several physically plausible structures reproduce the same measurement, should the goal be to represent that uncertainty or to select a low-defect structure? What additional physical assumptions justify preferring lower-entropy solutions? I can imagine, in other physical situations, may be selecting a solution that satisfies some known symmetries of the problem would be the right approach.

In many cases, the iso-g2g_2 solutions can be completely distinct structures whose “physical quality” can be difficult to evaluate without access to a reliable potential energy surface and simulation. For instance, ring size distributions can vary quite drastically at the same g2(r)g_2(r) (e.g., Fig. 2d), but are challenging to relate directly to an energy penalty. In many new systems, there is no clear prior that allows us to select one ring size distribution against another. The lack of symmetries in most amorphous structures also prevent us from hand-selecting what is the “right approach”. Therefore, in most real-world systems, it is challenging to distinguish between defects and physically relevant features in amorphous materials.

Our low-entropy criterion is a surrogate for this selection criterion. As mentioned before, it is well-known that multiple solutions can lead to the same measurement, and that degenerate solutions with high energy are much more likely than low-energy ones. Therefore, we posit that maximizing structural uniformity (even at the disordered limit) can relate to high-quality solutions. This uniformity is translated into a low information entropy solution, which is compatible with the theoretical framework of degeneracy shown in the manuscript.

Thus, to address the reviewer’s question about the main goal, it is much easier to obtain variable solutions at fixed measurement compared to sampling a physically meaningful structure. Indeed, if physical constraints are known beforehand, these may be useful in selecting for structures. However, introducing hand-crafted constraints somewhat ignores the main advantages of generative modeling, which learn a prior from the data rather than relying on an external prior where the data representation is illustrated.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.