Pith. sign in

REVIEW 3 major objections 5 minor 45 references

Meta-neural Topology Optimization: Knowledge Infusion with Meta-learning

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper shows that a meta-learned initialization for a neural-network reparameterization of the density field lets topology optimization converge in fewer iterations—often fewer than conventional density-based methods—without hurting…

desk verdict Useful empirical result on strain-energy warm starts, but the meta-learning attribution doesn't survive the paper's own pretraining baseline. read the letter →

arxiv 2502.01830 v2 pith:OXRQABEU submitted 2025-02-03 cs.CE physics.comp-ph

Igor Kuszczak , Gawel Kus , Federico Bosi , Miguel A. Bessa This is my paper · ORCID
classification cs.CEphysics.comp-ph MSC 68T0774P15
keywords topologyoptimizationmeta-learningimplicitneuralrepresentationsreparameterizationcomplianceminimizationstrainenergyconditioningReptilecross-resolutiontransfer
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

This paper proposes meta-neural topology optimization, a way to give structural design optimizers a learned memory of past problems. Instead of starting each new design from a uniform density field, the method trains a neural network whose initial weights, after a meta-learning step, produce task-specific starting densities conditioned on the strain energy field of a uniform design. The paper claims these learned starting points let the optimizer converge in fewer finite-element iterations than standard neural topology optimization, and in many cases fewer than conventional density-based methods, without sacrificing final design quality. The key evidence is a set of 3,000 compliance-minimization tasks: meta-learned initializations were fastest in 57.6% of in-distribution tasks and 74.1% of cross-resolution tasks, with average iteration counts of 103.51 versus 110.97 for standard topology optimization and 149.47 for the neural baseline. It also reports that the meta-learner spontaneously rediscovers strain energy patterns as good initial guesses, which matches engineering intuition.

What carries the argument

The machinery is a SIREN neural network with ResNet-style skip connections that parameterizes the density field, with a meta-learned initialization found by Reptile. The network takes spatial coordinates and the strain energy field $E(x)$ of the uniform-density design as input and outputs element densities, so one set of weights can generate different initial designs for different boundary conditions. The inner loop performs 10 topology-optimization steps on sampled tasks; the outer loop moves the initialization toward the average of the adapted weight vectors. Conditioning on the strain energy field (Eq. 7 post-processing) is what carries the task-specific information, and the authors show that removing it sharply reduces the gains.

What would settle it

Construct two families of tasks with identical log-normalized strain energy fields but different optimal layouts—for example, symmetric boundary conditions that produce the same field but require mirrored structures—and check whether the meta-learned initialization still converges faster than a uniform one. If the gains vanish for tasks that share the same conditioning field yet have distinct optima, the claim that the strain energy field is a sufficient conditioning signal would be refuted. Alternatively, one could stress-test cross-resolution transfer on a mesh 16x finer than training; the paper only demonstrates 4x.

Watch

Extended reading notes

Core claim

The central discovery is that a few-shot meta-learning algorithm—Reptile—can distill reusable design knowledge from partial optimization trajectories, so that a single set of neural-network weights encodes a family of initial designs indexed by boundary conditions and volume fraction. The network maps element coordinates and the log-normalized strain energy field of a uniform-density reference design to initial densities, and the meta-trained initialization adapts to a new task in ten gradient steps. The paper shows that this initialization accelerates convergence across in-distribution, out-of-distribution, and cross-resolution tests, and that initializations trained on 64x64 meshes transfer to 256x256 discretizations. A notable secondary finding is that directly pretraining the network to reproduce the strain energy field—rather than meta-learning—also yields strong initializations, which the authors interpret as evidence that the strain energy field itself is the robust prior that meta-learning discovers.

Load-bearing premise

The strain energy field of the uniform-density design, after log normalization, carries enough information about the task that a single network conditioned on it can produce effective starting densities for unseen loads, supports, volume fractions, and mesh sizes.

Editorial extensions

If this is right

  • Engineers facing repeated design problems with varying loads and supports can reuse a single learned initialization instead of restarting from a blank canvas.
  • Initializations trained on coarse meshes transfer to finer meshes, so meta-training cost can be paid once at low resolution and amortized over high-resolution production runs.
  • The method works with conventional density-based optimizers as well: initializing standard topology optimization with meta-learned, strain-energy-conditioned designs also speeds up convergence (Appendix F).
  • Because meta-training uses only partial optimization trajectories, it does not require large datasets of pre-optimized designs, avoiding biases introduced by the optimizer used to generate such datasets.

Reading between the lines

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

  • The Appendix D result—that a simple supervised pretraining to reproduce the strain energy field beats meta-learning in-distribution—suggests the meta-learner is rediscovering a known physical prior; a direct implication is that the marginal value of meta-learning over good physics-based initialization may be modest for compliance minimization, and the method's advantage may lie mostly in domains w
  • Meta-learning's discovered initial designs mirror filtered strain energy fields; one testable extension is to use the same bilevel framework with a different conditioning field (e.g., sensitivity maps or stress fields) for problems where strain energy is not the natural prior, such as stress-constrained or buckling design.
  • The worst-case 149.8% compliance deterioration at low volume fractions indicates that the learned initialization can bias the optimizer into poor local minima; a robustifying variant might meta-learn an ensemble of initializations or adaptively fall back to a uniform start when the volume fraction is low.
  • Cross-resolution transfer at 4x suggests the mesh-agnostic parameterization, not the specific initialization, is what enables the speedup; testing at 16x would clarify whether the benefit is the learned density pattern or the resolution-independent representation itself.
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 / 5 minor

Summary. The paper proposes meta-neural topology optimization, a method that uses Reptile-style meta-learning to learn an initialization for a SIREN-based neural reparameterization of density fields in compliance-minimization topology optimization. The network is conditioned on the log-normalized strain energy field of a uniform-density design (Eq. 1 and Appendix B.1), and the meta-learned initialization is evaluated on 3000 held-out tasks spanning in-distribution, out-of-distribution, and cross-resolution (64x64 to 256x256) settings. The main results are that the meta-learned initialization converges fastest in 57.6% of in-distribution and 74.1% of cross-resolution tasks, with average iteration counts of 103.51 and 100.90, respectively, versus 110.97 and 130.03 for standard TO and 149.47 and 152.04 for neural TO. Final thresholded design compliance is within a 5% tolerance of standard TO in all settings. The paper also reports that the learned initializations closely resemble filtered strain energy fields, and Appendix D shows that a simple pretrained network regressing the same strain energy field performs comparably or better in-distribution.

Significance. If the central claim holds, the paper would make a useful contribution: it shows that learned initializations can close the iteration-count gap between neural TO and conventional TO, and it demonstrates nontrivial cross-resolution transfer, which is valuable because it avoids retraining on fine meshes. The evaluation is unusually thorough for the field, with 3000 test tasks, performance profiles, thresholded and continuous compliance comparisons, and several ablations (including a no-conditioning control). The authors are also honest about limitations, including the worst-case degradation at low volume fractions and the computational cost of meta-training. However, the significance is materially reduced by the paper's own Appendix D, which shows that a cheap supervised pretraining baseline without any meta-learning achieves comparable or superior in-distribution performance. Unless the pretraining baseline is shown to fail in the cross-resolution or out-of-distribution settings, the paper's distinctive claim about meta-learning as the mechanism of knowledge transfer is not established.

major comments (3)
  1. [Initial designs / Appendix D] The load-bearing attribution to meta-learning is not isolated. Appendix D reports that a network pretrained with plain MSE regression to output the strain energy field E(x) requires the fewest iterations in 40.7% of in-distribution tasks versus 37.4% for meta-neural TO, and it also improves both continuous and thresholded compliance. Because this baseline has no inner loop and no task adaptation, the iteration savings attributed to Reptile can be obtained by a single supervised regression on the conditioning field alone. The main-text claim that 'meta-learned network initializations require fewer iterations' is therefore not specific to meta-learning. To support the stated contribution, the pretrained baseline must also be evaluated on the out-of-distribution and cross-resolution test sets; without that control, the headline cross-resolution result (74.1% fastest, 33.6% fewer iterations) may transfer as well, making the meta-learning machinery unnecessary for the paper's strongest result.
  2. [Appendix B.2] All iteration counts are based on a single stopping criterion, |L(θ_t) - L(θ_{t-1})| < ε(1+|L(θ_{t-1})|) with ε = 10^-5. This criterion is scale-dependent and may interact differently with the loss trajectories of the three methods, since neural TO and meta-neural TO use Adam while standard TO uses MMA. The reported averages (103.51 vs 110.97 vs 149.47) and the performance-profile percentages are therefore not shown to be robust to the choice of criterion or to stochastic variation across runs. No error bars, repeated runs, or alternative stopping criteria are reported. Given that the acceleration claim is the paper's central quantitative contribution, the authors should provide at least a sensitivity check with a different ε or a fixed-iteration comparison, and ideally repeated meta-training runs to assess variance.
  3. [Appendix E / Discussion] Appendix E shows that removing the strain-energy conditioning sharply reduces the meta-learning gains, and Appendix D shows that a network trained to regress the same strain-energy field performs as well as the meta-learned network. Taken together, these results suggest that the speedup is driven by the conditioning signal and by pretraining the network to output that signal, rather than by the bilevel meta-learning procedure. The discussion acknowledges this in a qualitative way, but the paper's framing and title emphasize meta-learning as the key knowledge-infusion mechanism. The authors should either (a) demonstrate a setting where the pretrained baseline fails but meta-neural TO succeeds (e.g., out-of-distribution or cross-resolution), or (b) reframe the contribution as a study of strain-energy-conditioned initialization for neural TO, with meta-learning presented as one of several ways to obtain such an initialization. As written, the central claim overstates the role of meta-learning.
minor comments (5)
  1. [References / Data availability] The abstract states that the method 'eliminates the need for pre-optimized training data,' but the pretraining baseline in Appendix D also uses no pre-optimized designs; it trains on strain-energy fields. Consider clarifying the distinction between the data used by the proposed method and that used by the baseline.
  2. [Results, Fig. 2] The text explains that performance profile percentages can sum to over 100% because of ties at the 200-iteration budget, but the caption and figure would benefit from explicitly noting this so readers do not misinterpret the fractions.
  3. [Appendix B.2] The hyperparameter search is described as 'focused grid search' but no grid ranges are reported. Please list the values considered for the inner and outer learning rates and ω0 so that the search is reproducible.
  4. [Appendix B.1] The normalization in Eq. (7) uses per-task min and max of log(E_raw). State explicitly whether the min/max are computed per task and how this behaves for near-uniform fields, which may be sensitive to numerical noise.
  5. [Data and code availability] The abstract and GitHub link state that code is available, but the 'Data and code availability' section says the datasets and code 'will be released upon publication.' Please reconcile these statements and clarify what is currently accessible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the meta-learned initialization is evaluated on held-out tasks, and the strain-energy 'discovery' is an ablation-supported observation rather than a fitted prediction.

full rationale

The claimed derivation chain is not circular. The meta-learned initialization theta is produced by Reptile on the normalized compliance loss L(theta, tau_i) = c(theta, tau_i) / c_ref(tau_i), and its speedup is measured on held-out in-distribution, out-of-distribution, and cross-resolution tasks against random-init neural TO and MMA; no parameter of the evaluation is fitted to those target results. The conditioning field E(x) is an input to the network by Eq. (1) and is computed once from a uniform-density reference design, not from the optimized design, so the faster convergence cannot be an artifact of encoding the answer. The paper's observation that meta-learned initial designs resemble filtered E(x) is an empirical property of the learned weights, and the authors actively test the alternative explanations: Appendix D shows a supervised identity-map pretraining also works, which weakens the attribution of the speedup to bilevel meta-learning but does not make the derivation circular, and Appendix E shows removing E(x) conditioning substantially reduces gains. The only self-citation, [12], supports the background claim that neural reparameterization has non-convex landscapes; this is independently corroborated by the paper's own neural TO baseline (149.47 average iterations in-distribution versus 103.51 for meta-neural TO). No equation-level reduction, fitted-parameter-as-prediction, or load-bearing self-citation chain is present. The over-attribution of the mechanism to meta-learning is a correctness or novelty concern, not a circularity one.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim is empirical and rests on the task-generation scheme, the neural reparameterization, and the Reptile update. There are no invented physical entities. The main free parameters are hyperparameters tuned on a 100-task validation set; the most consequential modeling assumption is that the uniform-design strain energy field is a sufficient conditioning signal, which the paper's Appendix D exposes as the real driver of performance.

free parameters (8)
  • inner learning rate = 1e-4
    Grid search on a 100-task held-out validation set; inner-loop Adam steps.
  • outer learning rate = 1e-6
    Grid search on validation set; Reptile outer-loop update.
  • SIREN frequency omega0 = 60.0 (meta), 30.0 (neural TO)
    Tuned per method on validation set; Appendix B.1.
  • inner loop steps = 10
    Chosen as a balance between computational efficiency and task-specific adaptation.
  • meta-batch size = 5
    Chosen so that 6000 iterations cover 30,000 training tasks exactly once.
  • meta-training iterations = 6000
    Set to avoid observed meta-overfitting beyond this point.
  • density filter radius = 1/32 domain size
    Standard density filter radius used in Appendix B.1.
  • sigmoid amplification factor = 10
    Chosen to improve contrast in the sigmoid-based filter; Appendix B.1.
assumptions (5)
  • domain assumption Compliance minimization with SIMP material interpolation is a valid model of the design problem (Eqs. 2-6).
    Adopted from refs [5,7]; standard in topology optimization, not proved here.
  • domain assumption A SIREN with skip connections can represent optimal topologies when driven by gradient descent.
    Based on refs [12,24,29]; required for neural reparameterization to work.
  • domain assumption The strain energy field of the uniform density design is a sufficient conditioning signal for all tasks.
    Central input in Eq. (1) and Appendix B.1; the whole method relies on this field being informative.
  • ad hoc to paper Reptile with 10 inner steps and small outer updates extracts transferable initialization knowledge.
    No theoretical guarantee; empirical choice validated only on the held-out validation set.
  • domain assumption The stopping criterion |L_t - L_{t-1}| < eps(1+|L_{t-1}|) with eps=1e-5 measures convergence fairly for all optimizers.
    Appendix B.2; iteration-count comparisons are only meaningful if this criterion is unbiased across methods.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Meta-neural Topology Optimization: Knowledge Infusion with Meta-learning." pith.science (2026). https://pith.science/paper/OXRQABEU

@misc{pith2026250201830,
  author       = {Pith},
  title        = {Pith review of: Meta-neural Topology Optimization: Knowledge Infusion with Meta-learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OXRQABEU}},
  note         = {Machine review of arXiv:2502.01830}
}
read the original abstract

When faced with novel design problems, traditional topology optimization methods discard all prior design experience and start from a uniform initial guess. While this avoids biasing the optimizer towards any particular solution, it also means that many computationally expensive iterations are needed to converge. Existing machine learning approaches address this through data-driven design prediction, but require large datasets of pre-optimized structures and often struggle to generalize across boundary conditions and mesh resolutions. We propose a new method, termed meta-neural topology optimization, which uses a meta-learning algorithm to learn effective initial designs for topology optimization with neural field parameterizations -- continuous, mesh-independent representations that encode material distributions in the weights of a neural network. Through bilevel optimization, our method distills reusable design knowledge from partial optimization trajectories, eliminating the need for pre-optimized training data. By conditioning the neural field on strain energy fields of reference designs, a single set of learned parameters encodes problem-specific initial structures for diverse boundary conditions. We evaluate our approach on 3000 compliance minimization tasks across in-distribution, out-of-distribution, and cross-resolution scenarios. Our method converges in fewer iterations in 57.6% of in-distribution and 74.1% of cross-resolution tasks, while maintaining design quality competitive with standard density-based topology optimization. Notably, initializations learned on coarse meshes transfer successfully to discretizations four times finer. Code is available at https://github.com/bessagroup/metatopia.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 32 canonical work pages

  1. [1]

    Albrechtsen, M. et al. Nanometer-scale photon confinement in topology-optimized dielectric cavities. Nature Communications 13, 6281 (2022)

    work page 2022

  2. [2]

    Müller, P. et al. Development of a density-based topology optimization of homogenized lattice structures for individualized hip endoprostheses and validation using micro-FE. Scientific Reports 14, 5719 (2024)

    work page 2024

  3. [3]

    Aage, N., Andreassen, E., Lazarov, B. S. & Sigmund, O. Giga-voxel computational morphogenesis for structural design. Nature 550, 84–86 (2017)

    work page 2017

  4. [4]

    Bendsøe, M. P. & Kikuchi, N. Generating optimal topologies in structural design using a homogenization method. Computer Methods in Applied Mechanics and Engineering 71, 197–224 (1988)

    work page 1988

  5. [5]

    & Maute, K

    Sigmund, O. & Maute, K. Topology optimization approaches. Structural and Multidisciplinary Optimization 48, 1031–1055 (2013)

    work page 2013

  6. [6]

    Bendsøe, M. P. Optimal shape design as a material distribution problem. Structural optimization 1, 193–202 (1989)

    work page 1989

  7. [7]

    A Class of Globally Convergent Optimization Methods Based on Conservative Convex Separable Approximations

    Svanberg, K. A Class of Globally Convergent Optimization Methods Based on Conservative Convex Separable Approximations. SIAM Journal on Optimization 12, 555–573 (2002)

    work page 2002

  8. [8]

    A., Choi, Y

    White, D. A., Choi, Y . & Kudo, J. A dual mesh method with adaptivity for stress-constrained topology optimization. Structural and Multidisciplinary Optimization 61, 749–762 (2020)

    work page 2020

Show all 45 references
  1. [9]

    & Greydanus, S

    Hoyer, S., Sohl-Dickstein, J. & Greydanus, S. Neural reparameterization improves structural optimization (2019). 1909.04240

    2019 arXiv

  2. [10]

    Deng, H. & To, A. C. Topology optimization based on deep representation learning (DRL) for compliance and stress-constrained design. Computational Mechanics 66, 449–469 (2020)

    2020

  3. [11]

    & Suresh, K

    Chandrasekhar, A. & Suresh, K. TOuNN: Topology Optimization using Neural Networks. Structural and Multidisciplinary Optimization 63, 1135–1149 (2021)

    2021

  4. [12]

    M., Aragon, A

    Sanu, S. M., Aragon, A. M. & Bessa, M. A. Neural topology optimization: The good, the bad, and the ugly (2024). 2407.13954

    2024

  5. [13]

    M., V ogl, C

    Herrmann, L., Sigmund, O., Li, V . M., V ogl, C. & Kollmannsberger, S. Neural Networks for Generating Better Local Optima in Topology Optimization (2024). 2407.17957

    2024 arXiv

  6. [14]

    & Jacot, A

    Dupuis, B. & Jacot, A. DNN-based Topology Optimisation: Spatial Invariance and Neural Tangent Kernel. In Advances in Neural Information Processing Systems, vol. 34, 27659–27669 (Curran Associates, Inc., 2021)

    2021

  7. [15]

    & Suresh, K

    Chandrasekhar, A. & Suresh, K. Length Scale Control in Topology Optimization using Fourier Enhanced Neural Networks (2021). 2109.01861

    2021 arXiv

  8. [16]

    & Babaei, V

    Doosti, N., Panetta, J. & Babaei, V . Topology Optimization via Frequency Tuning of Neural Design Representa- tions. In Symposium on Computational Fabrication, 1–9 (ACM, Virtual Event USA, 2021)

    2021

  9. [17]

    & Kara, L

    Nie, Z., Lin, T., Jiang, H. & Kara, L. B. TopologyGAN: Topology Optimization Using Generative Adversarial Networks Based on Physical Fields Over the Initial Domain. Journal of Mechanical Design 143, 031715 (2021)

    2021

  10. [18]

    & Chen, P

    Chen, H., Wu, R., Grinspun, E., Zheng, C. & Chen, P. Y . Implicit Neural Spatial Representations for Time- dependent PDEs. In Proceedings of the 40th International Conference on Machine Learning, 5162–5177 (PMLR, 2023)

    2023

  11. [19]

    H., Giannone, G., Regenwetter, L

    Nobari, A. H., Giannone, G., Regenwetter, L. & Ahmed, F. NITO: Neural Implicit Fields for Resolution-free Topology Optimization (2024). 2402.05073

    2024 arXiv

  12. [20]

    Zhang, Z. et al. TONR: An exploration for a novel way combining neural network with topology optimization. Computer Methods in Applied Mechanics and Engineering 386, 114083 (2021)

    2021

  13. [21]

    & Storkey, A

    Hospedales, T., Antoniou, A., Micaelli, P. & Storkey, A. Meta-Learning in Neural Networks: A Survey. IEEE transactions on pattern analysis and machine intelligence 44, 5149–5169 (2022). 8 A PREPRINT - FEBRUARY 5, 2025

    2022

  14. [22]

    & Levine, S

    Finn, C., Abbeel, P. & Levine, S. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks (2017). 1703.03400

    2017 arXiv

  15. [23]

    & Schulman, J

    Nichol, A., Achiam, J. & Schulman, J. On First-Order Meta-Learning Algorithms (2018). 1803.02999

    2018 arXiv

  16. [24]

    & Wetzstein, G

    Sitzmann, V ., Martel, J., Bergman, A., Lindell, D. & Wetzstein, G. Implicit Neural Representations with Periodic Activation Functions. In Advances in Neural Information Processing Systems , vol. 33, 7462–7473 (Curran Associates, Inc., 2020)

    2020

  17. [25]

    Tancik, M. et al. Learned Initializations for Optimizing Coordinate-Based Neural Representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2846–2855 (2021)

    2021

  18. [26]

    V ., Aage, N., Bærentzen, J

    Woldseth, R. V ., Aage, N., Bærentzen, J. A. & Sigmund, O. On the use of Artificial Neural Networks in Topology Optimisation. Structural and Multidisciplinary Optimization 65, 294 (2022). 2208.02563

    2022 arXiv

  19. [27]

    & Kang, N

    Shin, S., Shin, D. & Kang, N. Topology optimization via machine learning and deep learning: A review. Journal of Computational Design and Engineering 10, 1736–1766 (2023)

    2023

  20. [28]

    Lu, Y ., Jiang, K., Levine, J. A. & Berger, M. Compressive Neural Representations of V olumetric Scalar Fields. Computer Graphics Forum 40, 135–146 (2021)

    2021

  21. [29]

    & Thomaszewski, B

    Zehnder, J., Li, Y ., Coros, S. & Thomaszewski, B. NTopo: Mesh-free Topology Optimization using Implicit Neural Representations. In Advances in Neural Information Processing Systems, vol. 34, 10368–10381 (Curran Associates, Inc., 2021)

    2021

  22. [30]

    & Storkey, A

    Antoniou, A., Edwards, H. & Storkey, A. How to train your MAML (2019). 1810.09502

    2019 arXiv

  23. [31]

    Huisman, M., van Rijn, J. N. & Plaat, A. A Survey of Deep Meta-Learning. Artificial Intelligence Review 54, 4483–4541 (2021). 2010.03522

    2021 arXiv

  24. [32]

    Kingma, D. P. & Ba, J. Adam: A Method for Stochastic Optimization (2017). 1412.6980

    2017 arXiv

  25. [33]

    & Oseledets, I

    Sosnovik, I. & Oseledets, I. Neural networks for topology optimization (2017). 1709.09578

    2017 arXiv

  26. [34]

    & Ahmed, F

    Mazé, F. & Ahmed, F. Diffusion Models Beat GANs on Topology Optimization (2022). 2208.09591

    2022 arXiv

  27. [35]

    Johnson, S. G. The NLopt nonlinear-optimization package (2007)

    2007

  28. [36]

    On benchmarking and good scientific practise in topology optimization

    Sigmund, O. On benchmarking and good scientific practise in topology optimization. Structural and Multidisci- plinary Optimization 65, 315 (2022)

    2022

  29. [37]

    Dolan, E. D. & Moré, J. J. Benchmarking Optimization Software with Performance Profiles (2004). cs/0102001

    2004 arXiv

  30. [38]

    & Stolpe, M

    Rojas-Labanda, S. & Stolpe, M. Benchmarking optimization solvers for structural topology optimization. Structural and Multidisciplinary Optimization 52, 527–547 (2015)

    2015

  31. [39]

    Li, Z., Zhou, F., Chen, F. & Li, H. Meta-SGD: Learning to Learn Quickly for Few-Shot Learning (2017). 1707.09835

    2017 arXiv

  32. [40]

    E., Sarkar, S

    Lynch, M. E., Sarkar, S. & Maute, K. Machine Learning to Aid Tuning of Numerical Parameters in Topology Optimization. Journal of Mechanical Design 141 (2019)

    2019

  33. [41]

    & Burak Kara, L

    Chen, H., Joglekar, A. & Burak Kara, L. Topology Optimization Using Neural Networks With Conditioning Field Initialization for Improved Efficiency. In ASME 2023 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference (Amer...

    2023

  34. [42]

    Filters in topology optimization

    Bourdin, B. Filters in topology optimization. International Journal for Numerical Methods in Engineering 50, 2143–2158 (2001)

    2001

  35. [43]

    R., Tucker, R., Snavely, N

    Sitzmann, V ., Chan, E. R., Tucker, R., Snavely, N. & Wetzstein, G. MetaSDF: Meta-learning Signed Distance Functions (2020). 2006.09662

    2020 arXiv

  36. [44]

    Tancik, M. et al. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains (2020). 2006.10739

    2020 arXiv

  37. [45]

    Martins, J. R. R. A. & Ning, A. Engineering Design Optimization (Cambridge University Press, Cambridge, 2021). 9 A PREPRINT - FEBRUARY 5, 2025 A Problem formulation Compliance minimization. We use the following standard TO formulation to solve compliance minimization problems ...

    2021

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.