Pith. sign in

REVIEW 3 major objections 4 minor 78 references

Steering topology distributions for unified generative design of architected metamaterials

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that steering a pretrained diffusion model's topology distribution—rather than optimizing a single structure—makes one pretrained prior reusable across four metamaterial design tasks, including out-of-distribution targets.

desk verdict A useful engineering demonstration of steering a pretrained topology diffusion prior by iterative elite fine-tuning, with shipped code and honest limitations; the distribution-level claims outrun the evidence but the core idea is worth refereeing. read the letter →

arxiv 2607.24777 v1 pith:WCWNIHLF submitted 2026-06-15 cs.AI cond-mat.mtrl-scics.LG

Haolin Li , Yuyang Miao , Menglei Li , Jinshuai Bai , Liyuan Wang , Xin Liu , Bo Gao , Jiantao Liu
show 4 more authors
Danilo Mandic Zahra Sharif Khodaei M.H.Aliabadi Weiqiu Chen
This is my paper
classification cs.AIcond-mat.mtrl-scics.LG
keywords architectedmetamaterialsgenerativedesigntopologyoptimizationdiffusionmodelsstructuralpriorsdistributionsteeringauxetichomogenization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces Generative Topology Optimization (GenTO), which reframes metamaterial design as the task of steering a probability distribution over unit-cell topologies. GenTO pretrains a diffusion model on a dataset of synthetic topology families, then iteratively samples candidates, scores them with user-defined physical objectives, keeps the high-performing feasible elites, and fine-tunes the model on those elites. The authors claim that this loop induces an implicit distribution update that shifts probability mass toward task-specific high-performing regions while preserving diversity. They demonstrate the framework on four problems—thermal conductivity extremization, multi-objective morphology control, property-targeted auxetic design, and vibration transmission—where they report performance competitive with or better than task-specific optimizers, and in several cases generation reaching property space beyond the pretrained prior. The upshot, if true, is that a reusable topology prior plus a generic steering loop can replace task-specific conditional generators and hand-built optimization routines.

What carries the argument

The central mechanism is the distribution-steering loop: sample from the current diffusion model, binarize the signed-distance-function (SDF) fields at a prescribed volume fraction, screen by feasibility constraints, score via task-specific black-box evaluators (FFT homogenization for thermal and elastic properties, finite-element modal analysis for vibration), select elite samples, and fine-tune the diffusion model on those elite fields. The SDF representation is the key enabling substrate: it converts binary topologies into smooth continuous fields so that linear interpolation and cross-family mixing are meaningful, and it supports exact volume-fraction control via sample-wise quantile thr

What would settle it

Train GenTO on the same four synthetic families, but with an objective requiring a topology mode absent from the prior (e.g., a chiral or origami-like unit cell), and track the diversity of the elite set (e.g., pairwise Jaccard similarity or effective sample size) across 100 steering iterations. If the generated distribution collapses to near-duplicates of a handful of templates, or if the property-space envelope does not move beyond the convex hull of the pretrained prior for targets placed outside it, the distribution-steering and out-of-distribution claims would be falsified.

Watch

Extended reading notes

Core claim

GenTO's central discovery is that the object of optimization can be moved from a single topology to a task-adapted topology distribution. Starting from a diffusion model pretrained on four synthetic topology families (Gaussian Random Fields, Cahn-Hilliard Phase Fields, Porous Structures, and Truss-based Structures) under three symmetry groups, GenTO filters generated candidates by feasibility and objective, retains an elite set, and fine-tunes the model on those elite fields. The paper formalizes the update as p_{i+1} = U(p_i, p̃_i), interprets it as KL-regularized score dynamics whose β→∞ limit is hard thresholding, and argues that diffusion's stochastic kernels plus SDF-space cross-family

Load-bearing premise

The central load-bearing premise is that fine-tuning the diffusion model on a small elite set of high-scoring candidates will shift probability mass toward the objective region without mode collapse, and that the pretrained prior's coverage is rich enough for that shift to reach new task regions—both of which the paper's Discussion acknowledges as open.

Editorial extensions

If this is right

  • If the central claim holds, a single pretrained topology prior per symmetry class plus a generic steering loop can replace task-specific conditional generators and hand-coded optimization routines.
  • Because the objective enters only through sampled scores, the same framework applies to non-differentiable or black-box targets, such as thresholded transmission spectra, without requiring gradients.
  • Steering in distribution space yields not just one optimum but a family of high-performing designs, which is directly useful for multi-objective trade-offs and design-space exploration.
  • Reported out-of-distribution property matching suggests that the steering loop can push generation beyond the initial training data, expanding the reachable design space.
  • The modular split between offline prior learning and online steering implies that enriching the prior with new structure families benefits all downstream tasks without re-optimizing them.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If distribution steering preserves diversity as claimed, the same loop could transfer to other structured design spaces—photonic crystals, acoustic metamaterials, or composite microstructures—where a diffusion prior over geometry exists and evaluation is expensive.
  • The demonstration is confined to 2D unit cells and task-matched symmetry priors; a universal prior serving all tasks would require either a much larger dataset spanning many symmetry groups or a conditional symmetry mechanism, which the paper does not establish.
  • An explicit study of how diversity metrics (e.g., effective sample size or pairwise distance of elites) evolve over steering iterations would directly test the mode-collapse assumption; the paper reports best and mean scores but not a diversity metric on the final distribution.
  • The Gaussian-mixture interpretation of the elite set implies that the number of elite samples and the diffusion noise schedule jointly control the exploration/exploitation trade-off; an empirical study of this interaction could sharpen the method's operating point.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces Generative Topology Optimization (GenTO), a framework that pretrains a diffusion model on a dataset of binary unit-cell topologies converted to signed-distance fields, and then adapts this prior to specific design tasks through an iterative loop of sampling, feasibility screening, physics-based scoring, elite selection, and fine-tuning. The authors argue that this shifts optimization from a single topology to a task-adapted topology distribution, and demonstrate the approach on four benchmark problems: thermal conductivity extremization, multi-objective morphology control (feature size vs. fractal dimension), property-targeted auxetic design, and frequency-domain vibration transmission design. Claims include state-of-the-art or competitive performance versus topology optimization, genetic algorithms, best-in-prior, and best-in-training-set baselines, with quantitative scores reported for each case. The manuscript is accompanied by code/data availability statements and an extensive supplement containing solver details, baselines, and additional experiments.

Significance. If the central distribution-steering claim holds, GenTO offers a genuinely reusable design paradigm: one pretrained diffusion prior per symmetry class plus a generic steering loop handles scalar, multi-objective, property-targeted, and function-targeted design problems, including targets outside the initial property-space hull. Strengths include the use of standard, well-specified physics solvers (FFT homogenization, Galerkin-FFT elasticity, modal FEM), consistent constraint formulations across methods, fair evaluation-budget claims, and explicit code/data availability. The experimental validation of auxetic unit cells is a positive addition. However, the paper's headline conceptual contribution—that the method optimizes a topology distribution rather than a single structure—is not backed by direct distribution-level measurements; the existing evidence is largely score means, Pareto fronts, and qualitative galleries. The authors themselves concede in the Discussion that a rigorous theory of distribution steering is missing, and the supplement notes the framework is bounded by prior coverage. These gaps weaken, but do not invalidate, the empirical results.

major comments (3)
  1. [§4.1, Eq. (9); Supp. B.4, Algorithm 1] The steering update U is defined only as 'the implicit distribution update induced by fine-tuning the diffusion model on samples from p̃_i'. The paper never measures whether p_{i+1} is actually a broad distribution shift or a collapse onto a few memorized elite samples. No distribution divergence (e.g., MMD or Fréchet distance) between p_i and p_{i+1}, no nearest-neighbor diversity statistic on generated samples, no multi-seed variance, and no memorization test are reported. The Abstract's claim that GenTO 'preserves structural diversity' and Eq. (1)'s reliance on a task-adapted distribution p*_T are load-bearing; without such measurements, the method could reduce to a stochastic local search with retraining. The Discussion (§3) explicitly concedes the need for 'a more rigorous theory of distribution steering'. This is a central gap that can be addressed with additional experiments.
  2. [§4.3, 'Out-of-Distribution (OOD) Extrapolation'; Fig. 4e] The OOD claim is supported by showing that an auxetic target lies outside the convex hull of the pretrained distribution in property space, and by score distributions that extend beyond the initial range. Neither observation establishes that the generated topologies themselves are out-of-distribution in topology space. Property-space extrapolation can arise from combining existing morphological motifs without leaving the manifold of learned structures. The paper should quantify distribution shift in field space (e.g., distance from the training manifold, reconstruction error, or structural novelty metrics) and show example topologies that are not simple interpolations of training-family members. The Discussion's concession that OOD behavior lacks a rigorous theory makes this measurement gap especially important.
  3. [§4.3, Eqs. (20)–(23)] The theoretical interpretation introduces a KL-regularized optimization whose analytical solution is an exponentially tilted distribution q*_i, then states that in the β→∞ limit this degenerates to hard thresholding. However, no proof or empirical demonstration connects the actual fine-tuning update (Eq. 8) to the solution of Eq. (20). The GMM singularity-resolution argument is a plausible heuristic, but it does not establish that the fine-tuned model converges to q*_i or that it preserves diversity. The relationship between the algorithmic hyperparameters (τ_i, N_take, fine-tuning epochs, learning rate) and the conceptual parameters (β, the KL weight) is never specified. A formal or empirical mapping, even for a simplified setting, would strengthen the central distribution-steering claim considerably.
minor comments (4)
  1. [Supp. Contents vs. main text] Main text references 'Supplementary Section F.1' for equal evaluation budgets, but the Supplementary Information table of contents begins Extended Discussions at F.2; Section F.1 appears to be missing or mislabeled. This should be corrected.
  2. [Fig. 4e] The labels E_min, E_max, ν_xy,min, ν_xy,max are shown but not defined in either the main text or the caption. Please state the definition (e.g., empirical min/max over pretrained samples) and how the convex hull was computed.
  3. [Supp. Table 3] For thermal conductivity maximization, the Best in Prior baseline (0.2170) is worse than Best in training sets (0.2404). This is surprising given the prior is trained on the same data; a brief explanation (e.g., prior mode mixing, sampling variability, or low probability of high-conductivity samples) would help the reader interpret the baseline.
  4. [Eq. (39)] The diffusion loss is written without the expectation over t and ε, unlike Eq. (4). Adding the expectation or clarifying that Eq. (39) is the per-sample loss would avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: every reported design score is validated by independent FFT/FEM solvers against externally prescribed targets; the steering loop is the algorithm itself, not a fitted prediction.

full rationale

The central derivation chain is not circular. GenTO's outputs are evaluated by external physics solvers: thermal scores use FFT-based homogenization, auxetic tensors use Galerkin-FFT elasticity homogenization, and vibration scores use finite-element modal analysis. The optimization targets are prescribed by the user (e.g., target elasticity tensors in Eqs. 94-96, pass/gap frequency bands, thermal extremization direction), not derived from the generative model. The generative prior is pretrained on synthetic topology families, and the steering loop fine-tunes the diffusion generator on elite samples scored by those solvers; no fitted parameter is re-imported as a prediction. The paper's self-citations (Refs. [42], [72]) concern numerical homogenization routines and are not load-bearing for the distribution-steering claim. Section 4.3 is explicitly labeled an 'Interpretation of Distribution Steering,' and the paper itself concedes in the Discussion that 'A more rigorous theory of distribution steering is needed' — a limitation about theoretical justification, not evidence that the empirical results reduce to their inputs. The reported comparisons against Best-in-Prior, Best-in-training-sets, GA, and TO are genuine external benchmarks under the stated evaluation budgets. Thus no specific circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No invented physical entities: the 'task-adapted topology distribution' is a mathematical object (a weighted/truncated density), not a new degree of freedom, so invented_entities is empty. The honest ledger items are the hand-chosen steering constants (λ_κ, N_take, thresholds, LR/epochs) and the interpretive bridge between the KL-regularized optimum and the actual elite-fine-tuning procedure, which the paper itself describes as heuristic. The prior-coverage axiom is the paper's own stated limitation.

free parameters (6)
  • Anisotropy penalty weight λ_κ = 0.1 = 0.1
    Hand-chosen weight in the thermal steering score (Eq. 11). The reported benchmark quantity is the unpenalized mean conductivity, so the ranking used during selection differs from the reported metric; it biases both max and min solutions toward isotropy.
  • Per-case elite-set size N_take = 64 (C1), 2048 (C2), 256 (C3, C4)
    Controls selection pressure. For C2 (multi-objective), 2048 elites retained per iteration is a very large fraction of the 2×N_take validated samples and contributes to the breadth of the Pareto front.
  • Selection threshold τ_i (rank-based, with persistent elite pool) = top-N_take of merged Elite_{i-1} ∪ Valid_i
    Implements the β→∞ hard-step limit of Eq. (21). The elite pool persists across iterations (Eq. 43), an EMA-like smoothing choice not present in the theory.
  • Per-case fine-tuning hyperparameters = LR 1e-5..1e-4, epochs 10..100, batch 64
    Learning rate, epochs, and batch size jointly determine how far each fine-tune step moves the distribution; tuned per case (Supp Table 2).
  • Diffusion sampling stochasticity knob = not quoted
    Supp B.4: "an additional scalar parameter controls the magnitude of the stochastic term in the reverse step" — a free exploration-strength parameter whose value is not reported.
  • Vibration threshold T_th and band weights = -10 dB; 0.25/0.50/0.25
    Define the objective in Eq. (15) rather than being fitted to data; included for completeness since they set the difficulty of the reported 100% satisfaction.
assumptions (5)
  • domain assumption FFT-based periodic homogenization (Lippmann–Schwinger; Galerkin-FFT with MINRES) yields converged effective conductivity/elasticity tensors on 256×256 binary cells.
    All thermal and auxetic scores rest on this (Supp C.1, C.3); 500 fixed-point iterations / stated solver tolerance are assumed sufficient for the reported 0.04%–0.35% matching accuracy.
  • ad hoc to paper The four synthetic families (GRF, CHPF, PR, TS) under three symmetry classes constitute sufficient 'topology knowledge' for the four tasks.
    The prior is the paper's own procedurally generated dataset (Supp B.1); the paper concedes "the expressive range of the framework is still shaped by the coverage and richness of the pretrained topology prior" (Discussion).
  • standard math The KL-regularized distribution update (Eq. 20) with β→∞ reduces to threshold selection, and fine-tuning on the elite set (Eq. 8) implements that shift.
    This is standard reward-weighted regression in the RL/statistical-mechanics sense; the paper explicitly labels the implemented thresholding "the heuristic thresholding operation used in our framework" (§4.3), so the link is asserted, not derived.
  • domain assumption Gaussian reverse-diffusion kernels (Eqs. 24-25) and cross-family SDF overlap (Eqs. 26-28) provide the mechanism for out-of-distribution generation.
    Mathematically the kernels have full support, but that trivially makes every sample 'out-of-support'; the meaningful OOD evidence is the property-space hull test (Fig. 4e), not the kernel argument.
  • domain assumption The minimum-feature-size estimators (bridge width w_min and skeleton width w_skel, >5 px) are valid manufacturability proxies applied identically to all baselines.
    Supp C.1/C.3. The TO baseline's min(κ) 'failure' under this constraint shows the constraint materially shapes the comparison; its validity as a proxy is not independently established beyond the paper's geometric rationale.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Steering topology distributions for unified generative design of architected metamaterials." pith.science (2026). https://pith.science/paper/WCWNIHLF

@misc{pith2026260724777,
  author       = {Pith},
  title        = {Pith review of: Steering topology distributions for unified generative design of architected metamaterials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WCWNIHLF}},
  note         = {Machine review of arXiv:2607.24777}
}
read the original abstract

Architected metamaterials derive their functions from structure, creating vast opportunities to program physical responses through topology design. However, existing design methods are often tailored to individual design problems, making limited use of topology knowledge for effective and broadly applicable design as objectives, constraints, and physical functions change. Here we introduce Generative Topology Optimization (GenTO), a unified framework that turns a learned topology prior into a reusable design engine. GenTO trains a diffusion model on a large full-order topology dataset and then iteratively steers the resulting topology distribution toward task-specific high-performing regions using user-defined physical objectives and constraints. This shifts the object of optimization from a single structure to a task-adapted topology distribution. Across topology design problems spanning thermal extremization, multi-objective morphology control, property-targeted auxetic design, and vibration transmission design, GenTO reuses pretrained topology priors for heterogeneous tasks, preserves structural diversity, and reaches high-performing solutions supported by numerical benchmarks and experimental validation. These results establish reusable topology knowledge as a unified principle for effective and scalable architected metamaterial design.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

78 extracted references · 14 canonical work pages

  1. [1]

    Nature reviews physics1(3), 198–210 (2019)

    Kadic, M., Milton, G.W., Van Hecke, M., Wegener, M.: 3d metamaterials. Nature reviews physics1(3), 198–210 (2019)

    2019

  2. [2]

    Nature543(7646), 533–537 (2017)

    Berger, J., Wadley, H., McMeeking, R.: Mechanical metamaterials at the theoretical limit of isotropic elastic stiffness. Nature543(7646), 533–537 (2017)

    2017

  3. [3]

    Nature640(8060), 931–940 (2025)

    Zhao, T., Dang, X., Manos, K., Zang, S., Mandal, J., Chen, M., Paulino, G.H.: Modular chiral origami metamaterials. Nature640(8060), 931–940 (2025)

    2025

  4. [4]

    Nature639(8055), 639–645 (2025)

    Fang, X., Yu, D., Wen, J., Dai, Y., Begley, M.R., Gao, H., Gumbsch, P.: Large recoverable elastic energy in chiral metamaterials via twist buckling. Nature639(8055), 639–645 (2025)

    2025

  5. [5]

    Nature Reviews Materials2(11), 1–11 (2017)

    Bertoldi, K., Vitelli, V., Christensen, J., Van Hecke, M.: Flexible mechan- ical metamaterials. Nature Reviews Materials2(11), 1–11 (2017)

    2017

  6. [6]

    Current Opinion in Solid State and Materials Science24(5), 100853 (2020)

    Xin, L., Siyuan, Y., Harry, L., Minghui, L., Yanfeng, C.: Topological mechanical metamaterials: A brief review. Current Opinion in Solid State and Materials Science24(5), 100853 (2020)

    2020

  7. [7]

    Nature Physics14(8), 811–815 (2018)

    Liu, B., Silverberg, J.L., Evans, A.A., Santangelo, C.D., Lang, R.J., Hull, T.C., Cohen, I.: Topological kinematics of origami metamaterials. Nature Physics14(8), 811–815 (2018)

    2018

  8. [8]

    Progress in Materials Science94, 114–173 (2018)

    Yu, X., Zhou, J., Liang, H., Jiang, Z., Wu, L.: Mechanical metamateri- als associated with stiffness, rigidity and compressibility: A brief review. Progress in Materials Science94, 114–173 (2018)

    2018

Show all 78 references
  1. [9]

    Mechanics of Materials155, 103742 (2021)

    Zheng, Y., Wang, Y., Lu, X., Zheng, J., Qu, J.: Topology optimisation for isotropic mechanical metamaterials considering material uncertainties. Mechanics of Materials155, 103742 (2021)

    2021

  2. [10]

    Structural and Multidisciplinary Optimization41(2), 163–177 (2010)

    Diaz, A.R., Sigmund, O.: A topology optimization method for design of negative permeability metamaterials. Structural and Multidisciplinary Optimization41(2), 163–177 (2010)

    2010

  3. [11]

    Advanced Theory and Simulations2(7), 1900017 (2019)

    Li, W., Meng, F., Chen, Y., Li, Y.f., Huang, X.: Topology optimization of photonic and phononic crystals and metamaterials: a review. Advanced Theory and Simulations2(7), 1900017 (2019)

    2019

  4. [12]

    Structural and multidisciplinary optimization21(2), 120–127 (2001)

    Sigmund, O.: A 99 line topology optimization code written in matlab. Structural and multidisciplinary optimization21(2), 120–127 (2001)

    2001

  5. [13]

    Engineering Structures 322, 119194 (2025)

    Jeong, H., Batuwatta-Gamage, C., Bai, J., Rathnayaka, C., Zhou, Y., Unified generative design of architected metamaterials27 Gu, Y.: An advanced physics-informed neural network-based framework for nonlinear and complex topology optimization. Engineering Structures 322, 119194 (2025)

    2025

  6. [14]

    Computer Methods in Applied Mechanics and Engineering417, 116401 (2023)

    Jeong, H., Batuwatta-Gamage, C., Bai, J., Xie, Y.M., Rathnayaka, C., Zhou, Y., Gu, Y.: A complete physics-informed neural network-based framework for structural topology optimization. Computer Methods in Applied Mechanics and Engineering417, 116401 (2023)

    2023

  7. [15]

    Engineering Structures 278, 115484 (2023)

    Jeong, H., Bai, J., Batuwatta-Gamage, C.P., Rathnayaka, C., Zhou, Y., Gu, Y.: A physics-informed neural network-based topology optimization (pinnto) framework for structural optimization. Engineering Structures 278, 115484 (2023)

    2023

  8. [16]

    Composite Structures282, 115035 (2022)

    Dong, J., Hu, C., Holmes, J., Qin, Q.-H., Xiao, Y.: Structural optimisation of cross-chiral metamaterial structures via genetic algorithm. Composite Structures282, 115035 (2022)

    2022

  9. [17]

    Oxford Open Materials Science4(1), 001 (2024)

    Cerniauskas, G., Sadia, H., Alam, P.: Machine intelligence in meta- materials design: a review. Oxford Open Materials Science4(1), 001 (2024)

    2024

  10. [18]

    Composite Structures267, 113915 (2021)

    Wang, L., Liu, H.-T.: Parameter optimization of bidirectional re-entrant auxetic honeycomb metamaterial based on genetic algorithm. Composite Structures267, 113915 (2021)

    2021

  11. [19]

    Thin-Walled Structures, 114719 (2026)

    Ji, Z., Li, D., Zhai, X., Gao, J., Tang, Y., Dai, N., Liao, W., Zhao, Y., Xie, Y.M.: Designing anisotropic mechanical metamaterials and metastructures: A review. Thin-Walled Structures, 114719 (2026)

    2026

  12. [20]

    arXiv preprint arXiv:2410.02017 (2024)

    Fotopoulos, G.B., Popovich, P., Papadopoulos, N.H.: Review non- convex optimization method for machine learning. arXiv preprint arXiv:2410.02017 (2024)

    2024 arXiv

  13. [21]

    Nature Reviews Physics7(6), 299–312 (2025)

    Chen, Y., Fleury, R., Seppecher, P., Hu, G., Wegener, M.: Nonlocal metamaterials and metasurfaces. Nature Reviews Physics7(6), 299–312 (2025)

    2025

  14. [22]

    Nature Reviews Materials6(3), 207–225 (2021)

    Zangeneh-Nejad, F., Sounas, D.L., Al` u, A., Fleury, R.: Analogue com- puting with metamaterials. Nature Reviews Materials6(3), 207–225 (2021)

    2021

  15. [23]

    Nature Reviews Physics5(4), 218–235 (2023) 28Unified generative design of architected metamaterials

    Zhang, Z., Xu, L., Qu, T., Lei, M., Lin, Z.-K., Ouyang, X., Jiang, J.- H., Huang, J.: Diffusion metamaterials. Nature Reviews Physics5(4), 218–235 (2023) 28Unified generative design of architected metamaterials

    2023

  16. [24]

    Science Advances6(17) (2020)

    Mao, Y., He, Q., Zhao, X.: Designing complex architectured materi- als with generative adversarial networks. Science Advances6(17) (2020). https://doi.org/10.1126/sciadv.aaz4169

    2020 doi

  17. [25]

    Materi- als & Design211, 110178 (2021)

    Zheng, X., Chen, T.-T., Guo, X., Samitsu, S., Watanabe, I.: Controllable inverse design of auxetic metamaterials using deep learning. Materi- als & Design211, 110178 (2021). https://doi.org/10.1016/j.matdes.2021. 110178

    2021 doi

  18. [26]

    Materials & Design208, 109937 (2021)

    Challapalli, A., Patel, D., Li, G.: Inverse machine learning framework for optimizing lightweight metamaterials. Materials & Design208, 109937 (2021). https://doi.org/10.1016/j.matdes.2021.109937

    2021

  19. [27]

    Journal of Materials Research and Technology27, 6117–6134 (2023)

    Chiang, Y.-H., Tseng, B.-Y., Wang, J.-P., Chen, Y.-W., Tung, C.-C., Yu, C.-H., Chen, P.-Y., Chen, C.-S.: Generating three-dimensional bioinspired microstructures using transformer-based generative adversarial network. Journal of Materials Research and Technology27, 6117–6134 (...

    2023 doi

  20. [28]

    Nature Communications14(1) (2023)

    Zheng, L., Karapiperis, K., Kumar, S., Kochmann, D.M.: Unifying the design space and optimizing linear and nonlinear truss metamaterials by generative modeling. Nature Communications14(1) (2023). https://doi. org/10.1038/s41467-023-42068-x

    2023 doi

  21. [29]

    Nature Machine Intelligence5(12), 1466–1475 (2023)

    Bastek, J.-H., Kochmann, D.M.: Inverse design of nonlinear mechan- ical metamaterials via video denoising diffusion models. Nature Machine Intelligence5(12), 1466–1475 (2023). https://doi.org/10.1038/ s42256-023-00762-x

    2023

  22. [30]

    AIAA Journal62(7), 2433–2442 (2024)

    Liu, S., Acar, P.: Generative adversarial networks for inverse design of two- dimensional spinodoid metamaterials. AIAA Journal62(7), 2433–2442 (2024). https://doi.org/10.2514/1.J063697

    2024 doi

  23. [31]

    npj Computational Materials10(1) (2024)

    Kim, N., Lee, D., Kim, C., Lee, D., Hong, Y.: Simple arithmetic operation in latent space can generate a novel three-dimensional graph metamateri- als. npj Computational Materials10(1) (2024). https://doi.org/10.1038/ s41524-024-01430-3

    2024

  24. [32]

    Advanced Materials31(35) (2019)

    Ma, W., Cheng, F., Xu, Y., Wen, Q., Liu, Y.: Probabilistic represen- tation and inverse design of metamaterials based on a deep generative model with semi-supervised learning strategy. Advanced Materials31(35) (2019). https://doi.org/10.1002/adma.201901111

    2019 doi

  25. [33]

    Computer Methods in Applied Mechanics and Engineering372, 113377 (2020)

    Wang, L., Chan, Y.-C., Ahmed, F., Liu, Z., Zhu, P., Chen, W.: Deep gener- ative modeling for mechanistic-based learning and design of metamaterial systems. Computer Methods in Applied Mechanics and Engineering372, 113377 (2020). https://doi.org/10.1016/j.cma.2020.113377 Unifie...

    2020

  26. [34]

    Computer-Aided Design190, 103977 (2026)

    Du, L., Hu, J., Wang, S., Jiang, Y., Lei, N., He, Y., Luo, Z.: Topo- GenMeta: Generative design of metamaterials based on diffusion model with attention to topology. Computer-Aided Design190, 103977 (2026). https://doi.org/10.1016/j.cad.2025.103977

    2026

  27. [35]

    Smart Materials in Manufacturing4, 100129 (2026)

    Yang, Y., Wang, L., Zhai, X., Chen, K., Wu, W., Zhao, Y., Chen, F., Liu, L., Fu, X.-M.: Guided diffusion for fast inverse design of voxel-based mechanical metamaterials. Smart Materials in Manufacturing4, 100129 (2026). https://doi.org/10.1016/j.smmf.2026.100129

    2026

  28. [36]

    Engineering with Computers41(4), 2671–2687 (2025)

    Dedoncker, S., Donner, C., Bischof, R., Taenzer, L., Van Damme, B.: Generative inverse design of multimodal resonant structures for locally resonant metamaterials. Engineering with Computers41(4), 2671–2687 (2025). https://doi.org/10.1007/s00366-025-02130-2

    2025 doi

  29. [37]

    In: Volume 3A: 50th Design Automation Con- ference (DAC), IDETC-CIE2024

    Apaza, G., Selva, D.: Transfer learning in multi-objective generative design of metamaterials. In: Volume 3A: 50th Design Automation Con- ference (DAC), IDETC-CIE2024. American Society of Mechanical Engi- neers, ??? (2024). https://doi.org/10.1115/DETC2024-142346. American Soc...

    2024 doi

  30. [38]

    npj Metamaterials1(1) (2025)

    Abu-Mualla, M., Huang, J.: Inverse design of curved mechanical meta- materials with geometric AI: a generative diffusion operates in compact latent space of cellular structures. npj Metamaterials1(1) (2025). https: //doi.org/10.1038/s44455-025-00005-6

    2025 doi

  31. [39]

    Nature Communications17(1) (2026)

    Wang, L., Feng, J., Zhai, X., Han, J., Chen, K., Ma, W.W.S., Liu, L., Fu, X.-M.: Data-driven inverse design of multifunctional bicontinu- ous multiscale structures. Nature Communications17(1) (2026). https: //doi.org/10.1038/s41467-025-68089-2

    2026 doi

  32. [40]

    Xie, Z., Xu, W., Zhao, D., Zhang, W., Dong, D., Xu, B., Liu, N., Mao, S., Xue, T.: Physics-guided diffusion models for inverse design of disordered metamaterials (2026)

    2026

  33. [41]

    In: Proceedings of SIGGRAPH Conference Papers, pp

    Xue, T., Liu, L., Lu, L., Henderson, P., Tang, P., Li, H., Liu, J., Zhao, H., Peng, H., Bickel, B.: MIND: Microstructure INverse design with generative hybrid neural representation. In: Proceedings of SIGGRAPH Conference Papers, pp. 1–12. ACM, ??? (2025). https://doi.org/10.11...

    2025

  34. [42]

    Composite Structures, 119793 (2025)

    Li, M., Wang, B., Li, H.: An enhanced sub-clustering based self-consistent clustering analysis for efficient homogenization of thermal conductivity in woven composites. Composite Structures, 119793 (2025)

    2025

  35. [43]

    Acta Materialia85, 387–395 (2015)

    Hern´ andez-Nava, E., Smith, C., Derguti, F., Tammas-Williams, S., L´ eonard, F., Withers, P.J., Todd, I., Goodall, R.: The effect of density 30Unified generative design of architected metamaterials and feature size on mechanical properties of isostructural metallic foams prod...

    2015

  36. [44]

    Medical engineering & physics35(5), 576–582 (2013)

    Sanchez-Molina, D., Velazquez-Ameijide, J., Quintana, V., Arregui- Dalmases, C., Crandall, J.R., Subit, D., Kerrigan, J.R.: Fractal dimension and mechanical properties of human cortical bone. Medical engineering & physics35(5), 576–582 (2013)

    2013

  37. [45]

    Smart materials and structures27(2), 023001 (2018)

    Ren, X., Das, R., Tran, P., Ngo, T.D., Xie, Y.M.: Auxetic metamaterials and structures: a review. Smart materials and structures27(2), 023001 (2018)

    2018

  38. [46]

    Nature Reviews Materials1(3), 1–13 (2016)

    Cummer, S.A., Christensen, J., Al` u, A.: Controlling sound with acoustic metamaterials. Nature Reviews Materials1(3), 1–13 (2016)

    2016

  39. [47]

    Advanced Materials Technologies 7(6), 2100698 (2022)

    Gao, N., Zhang, Z., Deng, J., Guo, X., Cheng, B., Hou, H.: Acoustic meta- materials for noise reduction: a review. Advanced Materials Technologies 7(6), 2100698 (2022)

    2022

  40. [48]

    International Journal of Solids and Structures191, 293–306 (2020)

    An, X., Lai, C., Fan, H., Zhang, C.: 3d acoustic metamaterial-based mechanical metalattice structures for low-frequency and broadband vibra- tion attenuation. International Journal of Solids and Structures191, 293–306 (2020)

    2020

  41. [49]

    International Journal of Solids and Structures44(22-23), 7092–7109 (2007)

    de Kruijf, N., Zhou, S., Li, Q., Mai, Y.-W.: Topological design of structures and composite materials with multiobjectives. International Journal of Solids and Structures44(22-23), 7092–7109 (2007)

    2007

  42. [50]

    Advanced Materi- als31(34), 1803670 (2019)

    Yeo, S.J., Oh, M.J., Yoo, P.J.: Structurally controlled cellular architec- tures for high-performance ultra-lightweight materials. Advanced Materi- als31(34), 1803670 (2019)

    2019

  43. [51]

    Advanced Engineering Materials21(3), 1800864 (2019)

    Surjadi, J.U., Gao, L., Du, H., Li, X., Xiong, X., Fang, N.X., Lu, Y.: Mechanical metamaterials and their engineering applications. Advanced Engineering Materials21(3), 1800864 (2019)

    2019

  44. [52]

    Science334(6058), 962–965 (2011)

    Schaedler, T.A., Jacobsen, A.J., Torrents, A., Sorensen, A.E., Lian, J., Greer, J.R., Valdevit, L., Carter, W.B.: Ultralight metallic micro- lattices. Science334(6058), 962–965 (2011). https://doi.org/10.1126/ science.1211649

    2011

  45. [53]

    Science344(6190), 1373–1377 (2014)

    Zheng, X., Lee, H., Weisgraber, T.H., Shusteff, M., DeOtte, J., Duoss, E.B., Kuntz, J.D., Biener, M.M., Ge, Q., Jackson, J.A., Kucheyev, S.O., Fang, N.X., Spadaccini, C.M.: Ultralight, ultrastiff mechanical metama- terials. Science344(6190), 1373–1377 (2014). https://doi.org/1...

    2014

  46. [54]

    Nature Materials4(7), 518–524 (2005)

    Hollister, S.J.: Porous scaffold design for tissue engineering. Nature Materials4(7), 518–524 (2005). https://doi.org/10.1038/nmat1421

    2005 doi

  47. [55]

    Biomaterials Science8(1), 18–38 (2020)

    Zadpoor, A.A.: Meta-biomaterials. Biomaterials Science8(1), 18–38 (2020). https://doi.org/10.1039/c9bm01247h

    2020 doi

  48. [56]

    Nature Reviews Materials3(12), 460–472 (2018)

    Assouar, B., Liang, B., Wu, Y., Li, Y., Cheng, J.-C., Jing, Y.: Acoustic metasurfaces. Nature Reviews Materials3(12), 460–472 (2018). https:// doi.org/10.1038/s41578-018-0061-4

    2018 doi

  49. [57]

    Nature Reviews Materials7(9), 683–701 (2022)

    Xia, X., Spadaccini, C.M., Greer, J.R.: Responsive materials architected in space and time. Nature Reviews Materials7(9), 683–701 (2022). https: //doi.org/10.1038/s41578-022-00450-z

    2022 doi

  50. [58]

    Nature Computational Science 4(8), 567–573 (2024)

    He, Q., Ferracin, S., Raney, J.R.: Programmable responsive metamaterials for mechanical computing and robotics. Nature Computational Science 4(8), 567–573 (2024). https://doi.org/10.1038/s43588-024-00673-w

    2024 doi

  51. [59]

    Nature596(7873), 583–589 (2021)

    Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ron- neberger, O., Tunyasuvunakool, K., Bates, R., ˇZ ´ ıdek, A., Potapenko, A., Bridgland, A., Meyer, C., Kohl, S.A.A., Ballard, A.J., Cowie, A., Romera- Paredes, B.,et al.: Highly accurate protein structure predict...

    2021

  52. [60]

    Nature630(8016), 493–500 (2024)

    Abramson, J., Adler, J., Dunger, J., Evans, R., Green, T., Pritzel, A., Ron- neberger, O., Willmore, L., Ballard, A.J., Bambrick, J., Bodenstein, S.W., Evans, D.A., Hung, C.-C., O’Neill, M., Reiman, D., Tunyasuvunakool, K.,et al.: Accurate structure prediction of biomolecular ...

    2024

  53. [61]

    ACS Cen- tral Science4(2), 268–276 (2018)

    G´ omez-Bombarelli, R., Wei, J.N., Duvenaud, D., Hern´ andez-Lobato, J.M., S´ anchez-Lengeling, B., Sheberla, D., Aguilera-Iparraguirre, J., Hirzel, T.D., Adams, R.P., Aspuru-Guzik, A.: Automatic chemical design using a data-driven continuous representation of molecules. ACS C...

    2018 doi

  54. [62]

    Nature620(7976), 1089–1100 (2023)

    Watson, J.L., Juergens, D., Bennett, N.R., Trippe, B.L., Yim, J., Eise- nach, H.E., Ahern, W., Borst, A.J., Ragotte, R.J., Milles, L.F., Wicky, B.I.M., Hanikel, N., Pellock, S.J., Courbet, A., Sheffler, W., Wang, J., et al.: De novo design of protein structure and function wit...

    2023

  55. [63]

    Nature624(7990), 80–85 (2023)

    Merchant, A., Batzner, S., Schoenholz, S.S., Aykol, M., Cheon, G., Cubuk, 32Unified generative design of architected metamaterials E.D.: Scaling deep learning for materials discovery. Nature624(7990), 80–85 (2023). https://doi.org/10.1038/s41586-023-06735-9

    2023 doi

  56. [64]

    Structural and Multidisciplinary Optimization53(1), 175–192 (2016)

    Zegard, T., Paulino, G.H.: Bridging topology optimization and addi- tive manufacturing. Structural and Multidisciplinary Optimization53(1), 175–192 (2016). https://doi.org/10.1007/s00158-015-1274-4

    2016 doi

  57. [65]

    Progress in Materials Science138, 101129 (2023)

    Bayat, M., Zinovieva, O., Ferrari, F., Ayas, C., Langelaar, M., Spangen- berg, J., Salajeghe, R., Poulios, K., Mohanty, S., Sigmund, O., Hattel, J.: Holistic computational design within additive manufacturing through topology optimization combined with multiphysics multi-scale...

    2023

  58. [66]

    Journal of computational physics79(1), 12–49 (1988)

    Osher, S., Sethian, J.A.: Fronts propagating with curvature-dependent speed: Algorithms based on hamilton-jacobi formulations. Journal of computational physics79(1), 12–49 (1988)

    1988

  59. [67]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  60. [68]

    Computer Methods in Applied Mechanics and Engineering71(2), 197–224 (1988)

    Bendsøe, M.P., Kikuchi, N.: Generating optimal topologies in structural design using a homogenization method. Computer Methods in Applied Mechanics and Engineering71(2), 197–224 (1988). https://doi.org/10. 1016/0045-7825(88)90086-2

    1988

  61. [69]

    University of Michigan Press, Ann Arbor (1975)

    Holland, J.H.: Adaptation in Natural and Artificial Systems. University of Michigan Press, Ann Arbor (1975)

    1975

  62. [70]

    Addison-Wesley, Reading, MA (1989)

    Goldberg, D.E.: Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley, Reading, MA (1989)

    1989

  63. [71]

    Cahn, J.W., Hilliard, J.E.: Free energy of a nonuniform system. i. interfa- cial free energy. The Journal of Chemical Physics28(2), 258–267 (1958). https://doi.org/10.1063/1.1744102

    1958 doi

  64. [72]

    Journal of Physics: Conference Series3072, 012017 (2025)

    Li, H., Li, M., Sharif Khodaei, Z., Aliabadi, M.H.: A batch-processing fft- based method for efficient homogenisation of periodic media using gpu. Journal of Physics: Conference Series3072, 012017 (2025)

    2025

  65. [73]

    physics Letters A141(8-9), 386–390 (1989)

    Liebovitch, L.S., Toth, T.: A fast algorithm to determine fractal dimen- sions by box counting. physics Letters A141(8-9), 386–390 (1989)

    1989

  66. [74]

    Computer Methods in Applied Mechanics and Engineering157(1–2), 69–94 (1998)

    Moulinec, H., Suquet, P.: A numerical method for computing the overall response of nonlinear composites with complex microstructure. Computer Methods in Applied Mechanics and Engineering157(1–2), 69–94 (1998). https://doi.org/10.1016/S0045-7825(97)00218-1 Unified generative de...

    1998 doi

  67. [75]

    International Journal for Numerical Methods in Engineering 111(10), 903–926 (2017)

    Zeman, J., de Geus, T.W., Vondˇ rejc, J., Peerlings, R.H., Geers, M.G.: A finite element perspective on nonlinear fft-based micromechanical sim- ulations. International Journal for Numerical Methods in Engineering 111(10), 903–926 (2017). https://doi.org/10.1002/nme.5481

    2017 doi

  68. [76]

    Applied Mathematical Sciences, vol

    Allaire, G.: Shape Optimization by the Homogenization Method. Applied Mathematical Sciences, vol. 146. Springer, New York (2002). https://doi. org/10.1007/978-1-4684-9286-6

    2002 doi

  69. [77]

    IEEE Transactions on Evolutionary Computation6(2), 182–197 (2002)

    Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A fast and elitist multi- objective genetic algorithm: Nsga-ii. IEEE Transactions on Evolutionary Computation6(2), 182–197 (2002). https://doi.org/10.1109/4235.996017 Steering topology distributions for unified generative design...

    2002

  70. [78]

    At iterationm, τ (m) κ (r) = (k(r)−k 0)e (m)(r),(58) ˆe(m+1)(ξ) =− ˆΓ 0 κ(ξ) ˆτ (m) κ (ξ),ξ̸=0,(59) and the zero mode is reset to ¯eα

    ,ξ̸=0.(55) In the discrete solver used here, a staggered-grid central-difference form is adopted, so the continuous wave vector is replaced by its discrete counterpart 56Unified generative design of architected metamaterials ˜ξ= ( ˜ξ1, ˜ξ2), which gives ˆΓ0 ij(˜ξ) = ˜ξi ˜ξj k0...

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.