acceptodds
Under review as a conference paper at ICLR 2027

MetaGround: Physics-Grounded Conditional Diffusion for Nonlinear Metamaterial Inverse Design

Abstract

Generative models for nonlinear metamaterial inverse design are typically confined to a single mechanical property, overlooking the physical coupling between the stress-strain (SS) response and volume fraction (VF). These generators rely on surrogate proxies to rapidly pre-screen candidate geometries, yet evaluating with proxies not validated against physical ground truth produces an optimistic bias that inflates apparent performance, and low-accuracy pre-screening in turn requires more inference candidates for compensation. We present MetaGround, a DBTL-style framework in which conditional diffusion designs SDF candidate geometries, FEM simulation builds physical ground truth, a surrogate proxy tests candidate quality, and residual-guided adaptation learns to recalibrate the proxy. Within this loop, MetaGround introduces multi-physics conditioning architectures for SS and VF, a VF extrapolation protocol probing the generalization boundary, FEM-supervised adaptation driven by residual-guided selection, and finally an independent test set sampled from the full design space to anchor evaluation in FEM ground truth. Experiments show that the cross-attention + FiLM configuration (B2-PD) outperforms the baseline and is the most robust under deep VF extrapolation; the independent test set systematically confirms the optimistic bias of the proxy, and residual-guided selection improves proxy accuracy, raising the downstream screening pass rate from 34.4% to 39.6% at a fixed FEM budget.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Conditional diffusion, nonlinear metamaterial, inverse design, SDF, FEM, surrogate proxy, DBTL loop, volume fraction extrapolation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Pyw6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the inverse design of nonlinear mechanical metamaterials is studied. The conditional diffusion model is used to generate a two-dimensional SDF structure that meets the requirements of the target stress-strain (SS) curve and volume fraction (VF), and the NOT surrogate model and ABAQUS FEM are combined for candidate screening and high fidelity verification. A DBTL (design build test learn) closed loop based on residual guidance is designed to calibrate the surrogate model.

Critical strengths
  1. The three conditional injection mechanisms of B1/B2/B3 are compared in detail, and the advantage of cross attention+film (B2) in decoupling multiple physical field signals is established.
  2. Through the systematic FEM verification of r0-r30, the generalized boundary of SDF diffusion model extrapolated by deep VF is quantitatively detected.
  3. The observation that the proxy prediction accuracy is inconsistent with the actual sorting performance deserves attention
Critical weaknesses

1. Conceptual Conflation: OOD Generalization Failure vs. Selection-Induced Optimism

The surge in proxy error from 0.0620.062 to 0.1740.174 might driven by the fact that the independent 1K test set features a significantly wider Volume Fraction (VF) range ([0.226, 0.613] vs. [0.262, 0.553]) and more complex topologies (as acknowledged in Appendix A.2.3, Table 3). This demonstrates the proxy model's poor Out-of-Distribution (OOD) generalization, rather than the traditional notion of "optimistic bias."

The true "optimistic bias" (or selection-induced optimism) is more accurately captured in Appendix Table 11: for the same batch of generated candidates, the proxy's predicted error (Etarget-NOT=0.029E_{\text{target-NOT}} = 0.029) is systematically much lower than the true FEM error (Etarget-FEM=0.087E_{\text{target-FEM}} = 0.087). This occurs because the proxy inherently "favors" or over-ranks the specific samples it happens to predict most accurately during the Top-4 screening process (a classic manifestation of Goodhart's Law or surrogate overoptimization).

By conflating "cross-distribution generalization failure" (0.0620.062 vs. 0.1740.174) with "in-distribution selection bias" (0.0290.029 vs. 0.0870.087) under the single umbrella term of "optimistic bias," and subsequently using the former as the primary motivation to introduce the SFV loop, the authors commit a clear conceptual conflation that undermines the logical rigor of their methodological motivation.


2. Insufficient Ablation on Conditioning Capacity and Missing Full-Pool Baseline for SFV

While the empirical results are valuable, many experiments only prove the effectiveness of the complete scheme but fail to isolate why the specific mechanisms emphasized by the authors are effective.

  • Parameter Mismatch in Conditioning Modules: The total parameter count of B2 is only ~12.6% larger than B3, but its conditioning module contains roughly 6.7× more parameters. Consequently, the performance drop in B3 cannot be solely attributed to the absence of cross-attention, but may partly stem from insufficient capacity in its conditioning module. A parameter-matched ablation is necessary to isolate architectural benefits from capacity stacking.
  • Missing Full-Pool Baseline and the "Budget Paradox": The SFV loop utilizes a pool of 3,675 FEM-labeled samples over up to 10 rounds of residual-guided cumulative fine-tuning. However, the paper lacks a direct comparison with a S2 full-pool fine-tuning baseline.
    • The Budget Paradox: Crucially, the calculation of residuals inherently requires ground-truth FEM labels. Since the paper operates on a pre-existing pool of 3,675 FEM-labeled samples, the full set of labels is effectively available from the start. Therefore, the standard active learning justification (saving expensive labeling budget) does not strictly apply in this offline evaluation setting.
    • Furthermore, the R10 (Under the recommended 400-samples-per-round configuration) endpoint in the text appears to have consumed the entire 3,675-sample pool. Thus, the performance gains of SFV might merely result from having more optimization steps or repeated sample access, rather than the residual-guided sampling strategy itself.
    • Suggestion: We strongly suggest adding a baseline where S2 is fine-tuned on the full 3,675 samples, under the same model architecture, loss function, and total computational budget (even convergence). This is crucial to isolate and verify the independent contribution of the residual-guided selection mechanism.

3. Misalignment Between Sample Selection Strategy, Optimization Objective, and Downstream Task

The proposed SFV framework selects training samples based on the NOT–FEM relative L2 prediction error and fine-tunes the surrogate using MSE with teacher anchoring. However, the primary role of the surrogate in this pipeline seems to be rank candidate structures, rather than merely predict their stress–strain (SS) curves accurately.

The authors' own results reveal a discrepancy between prediction accuracy and ranking performance. Although the S1 Adapter achieves a lower OOD prediction error than S2+SFV (0.1490.149 vs. 0.1630.163), it exhibits worse within-cell Spearman correlation (0.1120.112 vs. 0.2370.237) and Top-1 FEM-valid rate (34.7%34.7\% vs. 39.6%39.6\%). This partly indicates that minimizing pointwise prediction error does not necessarily lead to better candidate ranking.

This raises two critical questions regarding the design of the SFV loop:

  1. Why does SFV prioritize samples with large prediction residuals rather than those causing incorrect candidate rankings?
  2. Why is the surrogate trained solely with a pointwise regression objective (MSE) instead of explicitly optimizing for ranking quality?

Suggestion: I recommend including comparisons with ranking-error-based sample selection (e.g., selecting samples that cause the most within-cell ranking inversions) and MSE combined with a pairwise/listwise ranking loss (e.g., Margin Ranking Loss or ListNet), under matched training data and computational budgets. Evaluating prediction error, Spearman correlation, and Top-1 FEM-valid rate across these variants would clarify whether the current sample selection and training objectives are truly optimal for the downstream design-screening task.

Questions for authors

How to control pointwise error and pairwise error, as well as their roles and downstream impacts are interesting issues, but the author has not studied them in depth.

Reviewer_n6LZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposed a pipeline for topology generation and fast simulation based on target volume fraction (VF) and stress-strain (SS) curves. The method may lead to useful engineering products, but lacks new intellectual contributions. My specific comments are as follows:

  1. The key challenge to be addressed seems to be to reach SS and VF targets through generation. The solution seems to be B2 (cross-attention + FiLM). It is not clear to the reviewer why and how this architecture helps address the challenge in a mechanistic way. There is an entire analysis section missing.

  2. Can you characterize the scope of SS curves the generator is targeting? Does your generator learn to intentionally create topologies that use local large-deformations or even contacts to achieve a specific type of SS? Again, a first-principled analysis on such capability of the proposed generator would greatly improve the relevance and quality of the paper.

  3. Since the paper is more about inverse design, it would be great if the authors provide a specific design case study to highlight the unique capabilities. It is also necessary to compare with a standard solver. Abacus can do inverse design.

  4. The writing has a strong AI-generated flavor. Some of the technical terms, such as FiLM, are introduced before defined, making the paper hard to read.

Critical strengths

No critical strengths.

Critical weaknesses

The paper misses critical analysis and discussion on why and how the proposed architecture uniquely addresses their proposed challenge. The writing about the method is hard to follow. The reviewer is thus not convinced with merely empirical results presented.

Questions for authors

Please see paper summary.

Reviewer_8endOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a conditional diffusion model approach for inverse design of 2D meta materials. The diffusion model provides SDF based material shape representations which are evaluated using a neural surrogate model. The diffusion model is conditioned on the desired stress-strain response and the volume fraction. The surrogate model regresses the stress strain response of the generated material and serves as an efficient proxy of FEM validation. Since the surrogate model can make errors, choosing materials based on the model can be too optimistic and introduce bias. Thus the paper proposes to continue training the surrogate model on FEM validation results. The approach is evaluated in different variants and compared with its base variant that only conditions on stress strain response.

Critical strengths
  • The evaluation of conditioning the diffusion model on two inputs is insightful.
Critical weaknesses
  • Conditional diffusion is rather straightforward and there is not much novelty other than the evaluation of different conditioning strategies.
  • The paper is poorly written and hard to comprehend. Several terminology are not properly introduced and defined, e.g., "DBTL"/"optimistic bias" in the abstract, "extrapolation magnitude", "generator arms", "six-level FEM campaign", "deep VF extrapolation", "source geometry", "anchoring scope", "pass@4", "extant structural archetype", "effective boundary", "family's mechanical margin", "preset ratio", "denominator-amplification caveat", "mechanism control", "deeper reductions".
  • "multi-physics" typically refers to the simulation of different physical processes, but this is not the case for this approach. Please revise the notion.
  • What is \mathcal{V} in Eq 1 ? Please define. What do the pointed brackets mean ?
  • The original paper of NOT should be referenced in line 143.
  • Eq 5 seems to be rather parts of the evaluation protocol than of the method. Why these r levels ?
  • l. 198, G_\phi is undefined.
  • Eq 11, \hat{x}_0 is undefined.
  • Eq 12, \hat{VF} is undefined.
  • The auxiliary loss in Eq 12 does not seem to fit the standard diffusion framework in a sound way. Please derive it. Was ist used in all variants in the experiments?
  • Eq 13, the \mu_* are undefined. What is a row in this context ?
  • Related work should also discuss works on guided diffusion such as [*1], [*2].
  • The SDF generation does not guarantee that the SDFs are valid, e.g., obey the Eikonal constraint. This might lead to artifacts and errors in the generated materials. This issue should be evaluated.
  • The experimental evaluation should be focused on the main results. The corresponding tables and figures need to be included in the main text. The main text should also explain the methods for the main results in sufficient detail, so the approaches can be rated (for example, Sec 3.3 does not mention and detail the two "trainers"). Forward referencing into the appendix and using the material as essential parts in the main paper violates the page limit.
  • The evaluation tries to interpret the results and explain why the architectural choices lead to the results. This seems to be rather hypothetical as there are no proofs of these statements. They should be stated more carefully as hypotheses. (Examples: l. 333, "B2 isolates VF through FiLM,...", l. 344 ff. "B2's full modality separation keeps...", etc)
  • l. 466, please avoid the term "trustworthy" in this context, as the proposed method might not guarantee that.

[*1] Kreber et al. Guided Diffusion by Optimized Loss Functions on Relaxed Parameters for Inverse Material Design. TMLR 2026. [*2] Zampini et al. Training-free constrained generation with stable diffusion models. NeurIPS, 2025.

Questions for authors

Unfortunately, the paper is difficult to read and needs a major revision of the presentation. Also terminology needs to be clarified.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.