Pith. sign in

REVIEW 3 major objections 6 minor 96 references

NOTES shrinks design space 10x, beats baselines on photonics and structures

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-09 02:28 UTC pith:R23QQEIG

load-bearing objection Solid integration of PCA + DeepONet + CMA-ES for PDE-constrained inverse design. The framework is real and useful, but the 'consistently outperforms' claim is undercut by L-BFGS matching it on MBB with 10x fewer evaluations. the 3 major comments →

arxiv 2607.07682 v1 pith:R23QQEIG submitted 2026-07-08 cs.LG

Neural Operator-enabled Topology-informed Evolutionary Strategy for PDE-Constrained Optimization

Xiangming Huang , Guannan Zhang , Lu Lu , Rapha\"el Pestourie This is my paper
classification cs.LG
keywords designinversenotesoptimizationevolutionarydimensionalityneuralstrategy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces NOTES (Neural Operator-enabled Topology-informed Evolutionary Strategy), a two-stage framework for PDE-constrained inverse design. The core idea is to compress a high-dimensional design space (e.g., 256 binary pixels for a nanophotonic beam deflector, or 12,288 pixels for a structural beam) into a 25- or 60-dimensional latent space using PCA, then train a DeepONet neural operator to decode latent vectors back into full design topologies with built-in binarization constraints. CMA-ES then searches this compact latent space, with each candidate design evaluated by an actual PDE solver rather than a surrogate. The key claim is that this combination — physics-informed dimensionality reduction via training data, nonlinear decoding via DeepONet, and global derivative-free search via CMA-ES — produces higher-quality, lower-variance designs than direct CMA-ES, gradient-based topology optimization, generative models (GLOnet), or gradient-based optimization in the same latent space (L-BFGS + DeepONet). On nanophotonic beam deflectors governed by Maxwell's equations, NOTES consistently achieves over 95% deflection efficiency across multiple wavelength-angle pairs, including conditions where no high-efficiency training design existed. On the MBB structural beam benchmark, it achieves compliance of 246, a 1.6% improvement over L-BFGS topology optimization. The paper also shows the trained DeepONet transfers across operating conditions, resolutions, and volume constraints without retraining.

Core claim

The central finding is that decoupling topology learning from physics simulation creates a more effective optimization pipeline than embedding physics directly into the neural network's loss function. By training DeepONet exclusively on high-performance designs obtained from prior PDE-constrained optimization, the learned latent space is implicitly biased toward physically meaningful geometric features. This bias, combined with CMA-ES's ability to escape poor local optima in low-dimensional spaces, yields consistently better designs than gradient-based methods in the same latent space (which frequently converge to poor local optima) or direct evolutionary search in the full design space (suc

What carries the argument

DeepONet (a neural operator with branch and trunk sub-networks whose Hadamard product maps latent vectors and spatial coordinates to design topologies), PCA-based latent space (25 dimensions for 1D photonics, 60 for 2D structures), CMA-ES (derivative-free evolutionary optimizer operating in latent space), and a differentiable PDE solver (Meent for RCWA in photonics, finite-element for structural mechanics) used for objective evaluation.

Load-bearing premise

The fairness of the baseline comparisons rests on unequal hyperparameter settings: NOTES uses an initial CMA-ES step size of 1000 while the baseline CMA-ES uses 10, and the baseline CMA-ES is given four times more iterations than NOTES. The paper states these values result from hyperparameter tuning but does not provide systematic justification or sensitivity analysis for why these specific settings constitute a fair comparison.

What would settle it

If the baseline CMA-ES and L-BFGS + DeepONet were retuned with comparable initial step sizes and iteration budgets and closed the performance gap to within 1-2% of NOTES, the claimed advantage of coupling CMA-ES with the DeepONet latent space would be substantially weakened.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the latent-space approach generalizes, engineering teams could precompute a library of optimized designs for one set of operating conditions, train a neural decoder, and then rapidly discover high-performance designs for new conditions without retraining the decoder.
  • The finding that CMA-ES outperforms L-BFGS in the same latent space suggests that latent optimization landscapes retain significant nonconvexity, and the choice of global vs. local optimizer matters even after dimensionality reduction.
  • The transferability of the trained DeepONet across resolutions and volume fractions without retraining could reduce the computational cost of multi-fidelity design campaigns by orders of magnitude, particularly for fine meshes where generating training data is expensive.
  • The ablation showing that physics-informed training data is essential (random data yields compliance 735 vs. 246 on MBB) implies that the design manifold structure, not the neural architecture alone, is what makes the latent space searchable.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. This manuscript introduces NOTES (Neural Operator-enabled Topology-informed Evolutionary Strategy), a framework that combines PCA-based dimensionality reduction, a DeepONet nonlinear decoder, and CMA-ES for PDE-constrained inverse design. The method is evaluated on two applications: 1D nanophotonic beam deflectors (256 design variables reduced to 25 latent dimensions) and 2D MBB beam structural optimization (12,288 variables reduced to 60). The authors report that NOTES achieves over 95% deflection efficiency on photonics and compliance of 246 on MBB beams, outperforming several baselines including direct CMA-ES, topology optimization, GLOnet, and L-BFGS+DeepONet. The paper includes ablation studies on the decoder choice (DeepONet vs PCA), training data (physics-informed vs random), and optimizer choice (CMA-ES vs L-BFGS), as well as transferability experiments across resolutions and operating conditions.

Significance. The paper addresses a practically important problem—making evolutionary strategies tractable for high-dimensional PDE-constrained optimization—and the combination of neural operator representations with CMA-ES is a reasonable methodological contribution. The ablation studies are comprehensive and include informative comparisons (DeepONet vs PCA decoder, physics-informed vs random data, CMA-ES vs L-BFGS). The transferability demonstrations across resolutions and volume constraints without retraining are a notable strength. The framework is applied to two genuinely different PDE domains, which strengthens the generality claim. Code is stated to be available on GitHub, which supports reproducibility. However, the significance is tempered by the fact that on the MBB beam, the simpler L-BFGS+DeepONet baseline matches NOTES performance at 10x lower cost, and on the 1D photonics case, the PCA decoder performs comparably to DeepONet, raising questions about when each component of the framework is actually necessary.

major comments (3)
  1. Section VI.B and SI Section 2.7: The headline claim that NOTES 'consistently achieves' superior performance across both applications is not supported by the MBB beam results. L-BFGS+DeepONet achieves identical compliance (246) with approximately 10x fewer function evaluations (SI Section 2.7, Figure S16). The paper acknowledges this ('L-BFGS + DeepONet requires approximately 10 times fewer function evaluations, making it a highly competitive alternative') but does not reconcile it with the abstract's claim of consistent outperformance. The advantage of CMA-ES over L-BFGS is confined to the photonics case, where the latent landscape has more poor local optima (Section VI.C). This makes the central claim problem-dependent rather than universal. The abstract and conclusion should be revised to accurately scope the claim—for example, by stating that NOTES provides lower variance and more bin
  2. Section V.C: The baseline comparison between NOTES and direct CMA-ES uses different initial step sizes (σ₀=1000 for NOTES vs σ₀=10 for baseline CMA-ES) and the baseline receives four times more iterations. The paper states these values result from 'hyperparameter tuning' but provides no sensitivity analysis or justification for why these specific settings are fair. While the authors note that the PCA latent coordinates are not normalized (explaining the larger σ₀ for NOTES), the asymmetry in iteration budget and step size makes the comparison difficult to interpret. A sensitivity analysis varying σ₀ for both methods, or at minimum a justification for why the chosen settings represent each method at its best, would substantially strengthen the comparison.
  3. Section VI.D and SI Section 1.5 (Figure S6): On the 1D nanophotonic application, the PCA decoder combined with CMA-ES performs comparably to NOTES, with 'an average improvement of approximately 1%' in some cases. This means the DeepONet decoder—the central architectural contribution—is not necessary for one of the two applications. The paper attributes this to the simpler geometry and lower compression ratio, which is plausible, but it undermines the generality of the DeepONet contribution. The paper should more clearly delineate when the nonlinear decoder is needed (high compression ratio, complex geometry) versus when a linear decoder suffices, and the abstract should not imply that DeepONet is universally beneficial.
minor comments (6)
  1. Section V.C: The different initial sampling radii for baseline CMA-ES and NOTES (σ₀=10 vs σ₀=1000) should be tabulated alongside all other hyperparameters for both applications in a single location, rather than mentioned inline in prose.
  2. Table I: The table caption states that best mean, best maximum, and lowest standard deviation are highlighted with green, blue, and orange respectively, but in the text version no color coding is visible. Ensure the published version has clear visual distinction or explicit markers.
  3. Section VI.B: The claim 'NOTES demonstrates the best overall performance in most cases' could be quantified—specify how many of the 12 operating conditions in Table I have NOTES as the best mean.
  4. Figure 5: The out-of-distribution results at 530 nm note that 'efficiencies are overestimated because visible-light absorption in silicon is neglected.' This caveat should be stated in the main text abstract or conclusion as well, not only in the figure caption.
  5. SI Section 2.4: The observation of mode collapse ('most optimized designs look very similar') is mentioned but not analyzed. A brief discussion of whether this limits design diversity for practical applications would be helpful.
  6. Section IV.C, Eq. (8): The notation 1_N for an N-dimensional vector of ones is introduced but could be confused with the latent dimension n. Consider using a different symbol (e.g., 𝟙_N) or clarifying at first use.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive reading of our manuscript. The referee raises three substantive points regarding (1) the scoping of our headline claims given the MBB beam results, (2) fairness of the CMA-ES baseline comparison, and (3) the necessity of the DeepONet decoder on the 1D photonics application. We address each below.

read point-by-point responses

  1. Referee: Section VI.B and SI Section 2.7: The headline claim that NOTES 'consistently achieves' superior performance across both applications is not supported by the MBB beam results. L-BFGS+DeepONet achieves identical compliance (246) with approximately 10x fewer function evaluations. The paper acknowledges this but does not reconcile it with the abstract's claim of consistent outperformance. The advantage of CMA-ES over L-BFGS is confined to the photonics case. The abstract and conclusion should be revised to accurately scope the claim.

    Authors: The referee is correct. On the MBB beam, L-BFGS+DeepONet achieves the same best compliance as NOTES at roughly 10x lower computational cost, and we already acknowledge this in SI Section 2.7. The abstract's current phrasing—'consistently achieves over 95 percent efficiency, outperforming CMA-ES, topology optimization, and other baselines'—is scoped to the photonics application, but the overall framing of the paper does imply broader superiority. We agree this needs to be corrected. We will revise the abstract, introduction, and conclusion to explicitly state that NOTES provides lower variance and more consistent convergence across both applications, but that on the MBB beam, L-BFGS+DeepONet is a competitive alternative with lower cost. We will also add a sentence in the conclusion noting that the advantage of CMA-ES over L-BFGS is problem-dependent: it is most pronounced when the latent landscape contains many poor local optima (as in the photonics case, Section VI.C), and diminishes when the landscape is smoother (as in the MBB case). This is an honest characterization of our results. revision: yes

  2. Referee: Section V.C: The baseline comparison between NOTES and direct CMA-ES uses different initial step sizes (σ₀=1000 for NOTES vs σ₀=10 for baseline CMA-ES) and the baseline receives four times more iterations. No sensitivity analysis or justification for why these settings are fair is provided.

    Authors: We appreciate the referee raising this point. The asymmetry in σ₀ arises from the fact that the PCA latent coordinates for the photonics application are not normalized, so the NOTES latent space has a much larger scale than the raw 256-dimensional design space used by direct CMA-ES. We chose σ₀ for each method based on preliminary tuning to give each method its best chance. However, we agree that without a sensitivity analysis, the reader cannot verify fairness. We will address this in two ways. First, we will add a sensitivity analysis varying σ₀ for both methods (e.g., σ₀ ∈ {100, 500, 1000, 2000} for NOTES and σ₀ ∈ {1, 5, 10, 50} for direct CMA-ES) and show that the qualitative conclusion—NOTES converges faster and more consistently—holds across a range of settings. Second, we will add explicit justification in Section V.C explaining the scale difference and why the chosen σ₀ values represent each method operating near its best regime. We note that the four-fold iteration advantage given to direct CMA-ES actually favors the baseline, not NOTES, so this asymmetry works against our method rather than for it. We will make this point explicit in the revised text. revision: yes

  3. Referee: Section VI.D and SI Section 1.5: On the 1D nanophotonic application, the PCA decoder combined with CMA-ES performs comparably to NOTES, with an average improvement of approximately 1% in some cases. The DeepONet decoder—the central architectural contribution—is not necessary for one of the two applications. The paper should more clearly delineate when the nonlinear decoder is needed versus when a linear decoder suffices, and the abstract should not imply that DeepONet is universally beneficial.

    Authors: The referee's observation is accurate and we do not dispute it. On the 1D photonics case, the compression ratio is moderate (256→25, ~10x) and the geometry is one-dimensional, so PCA linear reconstruction retains sufficient information and the DeepONet decoder provides only marginal benefit. On the 2D MBB beam, the compression is far more aggressive (12,288→60, ~200x) and the geometry is more complex, and here the DeepONet decoder is essential—PCA+CMA-ES achieves compliance of ~394 versus 246 for NOTES under the same budget. We already present this contrast in Section VI.D and the Discussion, but we agree the framing can be sharpened. We will revise the abstract to avoid implying that DeepONet is universally beneficial, and we will add a clear statement in the introduction and conclusion delineating the regime where the nonlinear decoder is needed (high compression ratio, complex geometry) versus where a linear decoder suffices (low compression, simple geometry). We believe this actually strengthens the paper's contribution by providing actionable guidance on when each component of the framework is warranted. revision: yes

Circularity Check

0 steps flagged

No significant circularity: the framework's warm-start bias is transparently disclosed and the method demonstrably exceeds its training data.

full rationale

The paper's central derivation chain is not circular. The latent space is constructed from pre-optimized high-performance designs (Section V.B), and CMA-ES then searches this space to find new optima. While this creates a warm-start bias (the reader's concern), the paper is transparent about this and explicitly demonstrates that NOTES finds designs exceeding the training data: for (1100nm, 60°), the best training design achieves 90% efficiency while NOTES produces 95% (Section VI.E, Figure 3). The ablation study with randomly-sampled data (Section VI.G) further shows the performance advantage is not merely an artifact of dimensionality reduction. The self-citations (e.g., to DeepONet [1], DeepXDE [44]) are standard references to tools and prior methods, not load-bearing uniqueness theorems invoked to forbid alternatives. The unequal hyperparameter settings (σ₀=1000 vs σ₀=10) and the L-BFGS+DeepONet efficiency comparison on MBB are concerns about experimental fairness and claim calibration, not circularity in the derivation chain. The derivation is self-contained against external benchmarks (GLOnet, topology optimization, CMA-ES baselines).

Axiom & Free-Parameter Ledger

10 free parameters · 4 axioms · 0 invented entities

No new physical entities, particles, forces, or mathematical objects are introduced. The method composes existing tools (PCA, DeepONet, CMA-ES) without postulating new entities.

free parameters (10)
  • Latent dimension n = 25 (photonics), 60 (structural)
    Chosen to explain >90% (photonics) and >83% (structural) of PCA variance. Not tuned to the target result but selected by variance threshold.
  • Binarization regularization weight α = Not specified in main text
    Controls the contribution of L_bias in Eq. 9. Value not reported.
  • Sigmoid scale factor β = Doubled every 40,000 epochs (photonics)
    Controls binarization sharpness; schedule is hand-designed.
  • Sigmoid threshold η = Initialized to average of training data
    Initial value is data-derived; subsequent adaptation not fully specified.
  • CMA-ES population size = 20 (photonics), 30 (structural)
    Standard CMA-ES parameter; not systematically justified.
  • CMA-ES initial σ₀ = 1000 (NOTES), 10 (baseline CMA-ES)
    Stated as resulting from hyperparameter tuning; the large asymmetry between NOTES and baseline is not justified.
  • CMA-ES max iterations = 100 (photonics)
    Baseline CMA-ES uses 4x more iterations; fairness of this allocation is unclear.
  • ADAM learning rate = 1e-3
    Standard choice.
  • Training epochs = 400,000 (photonics), 30,000 (structural)
    Selected by convergence observation; large difference between applications not explained.
  • Network architecture = 3 layers × 60 neurons (photonics), 4 layers × 512 neurons (structural)
    Chosen for reconstruction quality; not systematically optimized.
axioms (4)
  • domain assumption PCA on high-performance designs captures meaningful geometric features that span the feasible design manifold
    Invoked in Section IV.B and supported by Figure 2D clustering, but the completeness of the PCA representation for unseen operating conditions is not proven.
  • standard math DeepONet's universal approximation theorem for operators applies to the mapping from latent vectors to binary topology designs
    Invoked implicitly via the use of DeepONet (Ref [1]); the theorem guarantees approximation capacity but not optimization landscape properties.
  • domain assumption CMA-ES is an appropriate optimizer for the learned latent space
    Section IV.D applies CMA-ES without analyzing the latent landscape properties (convexity, smoothness) that would justify the choice over gradient methods.
  • domain assumption High-performance designs under one set of operating conditions share transferable topological features useful for unseen conditions
    Central to the transferability claims in Section VI.E; partially validated by out-of-distribution results but the 530nm results are acknowledged as overestimated.

pith-pipeline@v1.1.0-glm · 21485 in / 4807 out tokens · 463098 ms · 2026-07-09T02:28:22.291301+00:00 · methodology

0 comments
read the original abstract

The inverse design of physical systems governed by partial differential equations is computationally demanding due to the high dimensionality and non-convexity of design spaces. Generative models for inverse design often lack robustness and transferability, whereas evolutionary strategies are robust but struggle in high-dimensional spaces. This paper introduces a Neural Operator-enabled Topology-informed Evolutionary Strategy (NOTES) that integrates dimensionality reduction, representation learning, and evolutionary optimization for efficient and transferable inverse design. NOTES couples a DeepONet-based neural operator with the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to perform global optimization in a compact latent space that encodes topology-aware priors while discovering high-performance designs for unseen operating conditions. Applied to nanophotonic beam-deflector inverse design governed by Maxwell's equations, NOTES reduces the design dimensionality from 256 to 25 and consistently achieves over 95 percent efficiency, outperforming CMA-ES, topology optimization, and other baselines. Applied to structural optimization, NOTES discovers designs that achieve compliance down to 246. By decoupling topology learning of a DeepONet from the governing physics in a PDE solver, NOTES provides a flexible and transferable framework for the inverse design of physical systems.

Figures

Figures reproduced from arXiv: 2607.07682 by Guannan Zhang, Lu Lu, Rapha\"el Pestourie, Xiangming Huang.

Figure 1
Figure 1. Figure 1: (A) Overview of the neural operator-enabled topology-informed [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: (A-F) qualitatively present six optimized metagrating patterns gen [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: This figure demonstrates the robustness of NOTES when the neural [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

96 extracted references · 96 canonical work pages · 16 internal anchors

  1. [1]

    2013 , publisher=

    The finite element method: a practical course , author=. 2013 , publisher=

    work page 2013

  2. [2]

    Fast approximate solvers for metamaterials design in electromagnetism

    Fast approximate solvers for metamaterials design in electromagnetism , author=. arXiv preprint arXiv:2503.20065 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    nature , volume=

    Deep learning , author=. nature , volume=. 2015 , publisher=

    work page 2015

  4. [4]

    Nature Reviews Physics , volume=

    Neural operators for accelerating scientific simulations and design , author=. Nature Reviews Physics , volume=. 2024 , publisher=

    work page 2024

  5. [5]

    SIAM Journal on Scientific Computing , volume=

    MIONet: Learning multiple-input operators via tensor product , author=. SIAM Journal on Scientific Computing , volume=. 2022 , publisher=

    work page 2022

  6. [6]

    Computer Methods in Applied Mechanics and Engineering , volume=

    A comprehensive and fair comparison of two neural operators (with practical extensions) based on fair data , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2022 , publisher=

    work page 2022

  7. [7]

    Nature Communications , volume=

    One-shot learning for solution operators of partial differential equations , author=. Nature Communications , volume=. 2025 , publisher=

    work page 2025

  8. [8]

    Solving forward and inverse PDE problems on unknown manifolds via physics-informed neural operators

    Solving forward and inverse pde problems on unknown manifolds via physics-informed neural operators , author=. arXiv preprint arXiv:2407.05477 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    Nature Computational Science , volume=

    A scalable framework for learning the geometry-dependent solution operators of partial differential equations , author=. Nature Computational Science , volume=. 2024 , publisher=

    work page 2024

  10. [10]

    Adam: A Method for Stochastic Optimization

    Adam: A method for stochastic optimization , author=. arXiv preprint arXiv:1412.6980 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Mathematical programming , volume=

    On the limited memory BFGS method for large scale optimization , author=. Mathematical programming , volume=. 1989 , publisher=

    work page 1989

  12. [12]

    Proceedings of the 2022 2nd International Conference on Control and Intelligent Robotics , pages=

    Principal Component Analysis-Improved Fuzzy Genetic Algorithm , author=. Proceedings of the 2022 2nd International Conference on Control and Intelligent Robotics , pages=

    work page 2022

  13. [13]

    Neural Computing and Applications , volume=

    Evolutionary algorithms and their applications to engineering problems , author=. Neural Computing and Applications , volume=. 2020 , publisher=

    work page 2020

  14. [14]

    1992 , publisher=

    Adaptation in natural and artificial systems: an introductory analysis with applications to biology, control, and artificial intelligence , author=. 1992 , publisher=

    work page 1992

  15. [15]

    Nature Machine Intelligence , volume=

    Designing neural networks through neuroevolution , author=. Nature Machine Intelligence , volume=. 2019 , publisher=

    work page 2019

  16. [16]

    The Journal of Machine Learning Research , volume=

    Natural evolution strategies , author=. The Journal of Machine Learning Research , volume=. 2014 , publisher=

    work page 2014

  17. [17]

    Swarm and evolutionary computation , volume=

    Recent advances in differential evolution--an updated survey , author=. Swarm and evolutionary computation , volume=. 2016 , publisher=

    work page 2016

  18. [18]

    Evolutionary computation , volume=

    Completely derandomized self-adaptation in evolution strategies , author=. Evolutionary computation , volume=. 2001 , publisher=

    work page 2001

  19. [19]

    Proceedings of the Genetic and Evolutionary Computation Conference 2016 , pages=

    Projection-based restricted covariance matrix adaptation for high dimension , author=. Proceedings of the Genetic and Evolutionary Computation Conference 2016 , pages=

    work page 2016

  20. [20]

    2005 IEEE congress on evolutionary computation , volume=

    A restart CMA evolution strategy with increasing population size , author=. 2005 IEEE congress on evolutionary computation , volume=. 2005 , organization=

    work page 2005

  21. [21]

    IEEE Transactions on Evolutionary Computation , volume=

    Large scale black-box optimization by limited-memory matrix adaptation , author=. IEEE Transactions on Evolutionary Computation , volume=. 2018 , publisher=

    work page 2018

  22. [22]

    The CMA Evolution Strategy: A Tutorial

    The CMA evolution strategy: A tutorial , author=. arXiv preprint arXiv:1604.00772 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    Swarm and Evolutionary Computation , volume=

    Large-scale evolutionary optimization: A review and comparative study , author=. Swarm and Evolutionary Computation , volume=. 2024 , publisher=

    work page 2024

  24. [24]

    Expert Systems with Applications , volume=

    A review of surrogate-assisted evolutionary algorithms for expensive optimization problems , author=. Expert Systems with Applications , volume=. 2023 , publisher=

    work page 2023

  25. [25]

    Information Sciences , volume=

    Decomposition-based evolutionary algorithm for large scale constrained problems , author=. Information Sciences , volume=. 2015 , publisher=

    work page 2015

  26. [26]

    Applied Soft Computing , volume=

    A clustering and dimensionality reduction based evolutionary algorithm for large-scale multi-objective problems , author=. Applied Soft Computing , volume=. 2020 , publisher=

    work page 2020

  27. [27]

    Journal of Petroleum Science and Engineering , volume=

    Surrogate-assisted evolutionary algorithm with dimensionality reduction method for water flooding production optimization , author=. Journal of Petroleum Science and Engineering , volume=. 2020 , publisher=

    work page 2020

  28. [28]

    IEEE Transactions on computers , volume=

    A nonlinear mapping for data structure analysis , author=. IEEE Transactions on computers , volume=. 2006 , publisher=

    work page 2006

  29. [29]

    Covariance Matrix Adaptation Evolution Strategy Assisted by Principal Component Analysis

    Covariance Matrix Adaptation Evolution Strategy Assisted by Principal Component Analysis , author=. arXiv preprint arXiv:2105.03687 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    Swarm and Evolutionary Computation , volume=

    Dimensionality reduction in evolutionary algorithms-based feature selection for motor imagery brain-computer interface , author=. Swarm and Evolutionary Computation , volume=. 2020 , publisher=

    work page 2020

  31. [31]

    2015 , publisher=

    Applied dynamic programming , author=. 2015 , publisher=

    work page 2015

  32. [32]

    Nature Reviews Physics , volume=

    Physics-informed machine learning , author=. Nature Reviews Physics , volume=. 2021 , publisher=

    work page 2021

  33. [33]

    Nature machine intelligence , volume=

    Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators , author=. Nature machine intelligence , volume=. 2021 , publisher=

    work page 2021

  34. [34]

    Efficient neural topology optimization via active learning for enhancing turbulent mass transfer in fluid channels

    Efficient neural topology optimization via active learning for enhancing turbulent mass transfer in fluid channels , author=. arXiv preprint arXiv:2503.03997 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [35]

    Nanophotonics , volume=

    Simulator-based training of generative neural networks for the inverse design of metasurfaces , author=. Nanophotonics , volume=. 2020 , publisher=

    work page 2020

  36. [36]

    Nano letters , volume=

    Global optimization of dielectric metasurfaces using a physics-driven neural network , author=. Nano letters , volume=. 2019 , publisher=

    work page 2019

  37. [37]

    ACS Photonics , volume=

    Design space reparameterization enforces hard geometric constraints in inverse-designed nanophotonic devices , author=. ACS Photonics , volume=. 2020 , publisher=

    work page 2020

  38. [38]

    arXiv preprint arXiv:2503.00199 , year=

    Seeded Topology Optimization for Commercial Foundry Integrated Photonics , author=. arXiv preprint arXiv:2503.00199 , year=

    work page arXiv

  39. [39]

    Nanophotonics , volume=

    Enhancing metasurface fabricability through minimum feature size enforcement , author=. Nanophotonics , volume=. 2024 , publisher=

    work page 2024

  40. [40]

    Journal of the Optical Society of America B , volume=

    Validation and characterization of algorithms and software for photonics inverse design , author=. Journal of the Optical Society of America B , volume=. 2024 , publisher=

    work page 2024

  41. [41]

    Optics express , volume=

    MetaNet: a new paradigm for data sharing in photonics research , author=. Optics express , volume=. 2020 , publisher=

    work page 2020

  42. [42]

    Nano letters , volume=

    Large-angle, multifunctional metagratings based on freeform multimode geometries , author=. Nano letters , volume=. 2017 , publisher=

    work page 2017

  43. [43]

    Optics express , volume=

    Inverse design of large-area metasurfaces , author=. Optics express , volume=. 2018 , publisher=

    work page 2018

  44. [44]

    npj Computational Materials , volume=

    Inverse design of metasurfaces with non-local interactions , author=. npj Computational Materials , volume=. 2020 , publisher=

    work page 2020

  45. [45]

    Optics express , volume=

    Genetically optimized all-dielectric metasurfaces , author=. Optics express , volume=. 2017 , publisher=

    work page 2017

  46. [46]

    Chip , pages=

    A Hybrid Algorithm-driven Approach for Efficient Design of Terahertz Molecule-Specific Metasensors , author=. Chip , pages=. 2025 , publisher=

    work page 2025

  47. [47]

    Optics Express , volume=

    High efficiency and large angle polarization independent beam deflection metagrating by Bayesian optimization , author=. Optics Express , volume=. 2025 , publisher=

    work page 2025

  48. [48]

    Meent: Differentiable Electromagnetic Simulator for Machine Learning

    Meent: Differentiable electromagnetic simulator for machine learning , author=. arXiv preprint arXiv:2406.12904 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  49. [49]

    Proceedings of the 12th annual conference companion on Genetic and evolutionary computation , pages=

    Comparing results of 31 algorithms from the black-box optimization benchmarking BBOB-2009 , author=. Proceedings of the 12th annual conference companion on Genetic and evolutionary computation , pages=

    work page 2009

  50. [50]

    Journal of Global Optimization , volume=

    Derivative-free optimization: a review of algorithms and comparison of software implementations , author=. Journal of Global Optimization , volume=. 2013 , publisher=

    work page 2013

  51. [51]

    0.0 , author=

    CMA-ES/pycma: r4. 0.0 , author=. 2024 , publisher=

    work page 2024

  52. [52]

    International conference on parallel problem solving from nature , pages=

    A simple modification in CMA-ES achieving linear time and space complexity , author=. International conference on parallel problem solving from nature , pages=. 2008 , organization=

    work page 2008

  53. [53]

    Proceedings of the 2014 Annual Conference on Genetic and Evolutionary Computation , pages=

    A computationally efficient limited memory CMA-ES for large scale optimization , author=. Proceedings of the 2014 Annual Conference on Genetic and Evolutionary Computation , pages=

    work page 2014

  54. [54]

    Uncertainty in Artificial Intelligence , pages=

    Enabling long-range exploration in minimization of multimodal functions , author=. Uncertainty in Artificial Intelligence , pages=. 2021 , organization=

    work page 2021

  55. [55]

    Covariance Matrix Adaptation Evolution Strategy for Low Effective Dimensionality

    Covariance Matrix Adaptation Evolution Strategy for Low Effective Dimensionality , author=. arXiv preprint arXiv:2412.01156 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  56. [56]

    Advancing CMA-ES with Learning-Based Cooperative Coevolution for Scalable Optimization

    Advancing cma-es with learning-based cooperative coevolution for scalable optimization , author=. arXiv preprint arXiv:2504.17578 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  57. [57]

    Neural reparameterization improves structural optimization

    Neural reparameterization improves structural optimization , author=. arXiv preprint arXiv:1909.04240 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1909

  58. [58]

    Generative Neural Reparameterization for Differentiable PDE-constrained Optimization

    Generative neural reparameterization for differentiable PDE-constrained optimization , author=. arXiv preprint arXiv:2410.12683 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  59. [59]

    arXiv preprint arXiv:2507.00546 , year=

    Inverse Design in Nanophotonics via Representation Learning , author=. arXiv preprint arXiv:2507.00546 , year=

    work page arXiv

  60. [60]

    Raissi and P

    M. Raissi and P. Perdikaris and G.E. Karniadakis , keywords =. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , journal =. 2019 , issn =. doi:https://doi.org/10.1016/j.jcp.2018.10.045 , url =

    work page doi:10.1016/j.jcp.2018.10.045 2019

  61. [61]

    Optics express , volume=

    Physics-informed neural networks for inverse problems in nano-optics and metamaterials , author=. Optics express , volume=. 2020 , publisher=

    work page 2020

  62. [62]

    SIAM Journal on Scientific Computing , volume=

    Physics-informed neural networks with hard constraints for inverse design , author=. SIAM Journal on Scientific Computing , volume=. 2021 , publisher=

    work page 2021

  63. [63]

    ACS Photonics , volume=

    High speed simulation and freeform optimization of nanophotonic devices with physics-augmented deep learning , author=. ACS Photonics , volume=. 2022 , publisher=

    work page 2022

  64. [64]

    Communications Engineering , volume=

    Large area optimization of meta-lens via data-free machine learning , author=. Communications Engineering , volume=. 2023 , publisher=

    work page 2023

  65. [65]

    Advanced Materials , volume=

    Probabilistic representation and inverse design of metamaterials based on a deep generative model with semi-supervised learning strategy , author=. Advanced Materials , volume=. 2019 , publisher=

    work page 2019

  66. [66]

    Nano letters , volume=

    Generative model for the inverse design of metasurfaces , author=. Nano letters , volume=. 2018 , publisher=

    work page 2018

  67. [67]

    ACS photonics , volume=

    Inverse design of unitary transmission matrices in silicon photonic coupled waveguide arrays using a neural adjoint model , author=. ACS photonics , volume=. 2025 , publisher=

    work page 2025

  68. [68]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Representation learning: A review and new perspectives , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2013 , publisher=

    work page 2013

  69. [69]

    Digital Discovery , volume=

    Deep-learning enabled photonic nanostructure discovery in arbitrarily large shape sets via linked latent space representation learning , author=. Digital Discovery , volume=. 2024 , publisher=

    work page 2024

  70. [70]

    Electronics , volume=

    An inverse design framework for isotropic metasurfaces based on representation learning , author=. Electronics , volume=. 2022 , publisher=

    work page 2022

  71. [71]

    Engineering Applications of Artificial Intelligence , volume=

    Inverse design of non-parametric acoustic metamaterials via transfer-learned dual variational autoencoder with latent space-based data augmentation , author=. Engineering Applications of Artificial Intelligence , volume=. 2025 , publisher=

    work page 2025

  72. [72]

    Optics Express , volume=

    Active learning for efficient nanophotonics inverse design in large and diverse design spaces , author=. Optics Express , volume=. 2025 , publisher=

    work page 2025

  73. [73]

    Matter , volume=

    An invertible crystallographic representation for general inverse design of inorganic crystals with targeted properties , author=. Matter , volume=. 2022 , publisher=

    work page 2022

  74. [74]

    Nature Communications , volume=

    An invertible, invariant crystal representation for inverse design of solid-state materials using generative deep learning , author=. Nature Communications , volume=. 2023 , publisher=

    work page 2023

  75. [75]

    Small Methods , pages=

    HiLAB: A Hybrid Inverse-Design Framework , author=. Small Methods , pages=. 2025 , publisher=

    work page 2025

  76. [76]

    On the use of Artificial Neural Networks in Topology Optimisation

    On the use of artificial neural networks in topology optimisation , author=. arXiv preprint arXiv:2208.02563 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  77. [77]

    On Incorporating Inductive Biases into VAEs

    On incorporating inductive biases into vaes , author=. arXiv preprint arXiv:2106.13746 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  78. [78]

    Failure Modes of Variational Autoencoders and Their Effects on Downstream Tasks

    Failure modes of variational autoencoders and their effects on downstream tasks , author=. arXiv preprint arXiv:2007.07124 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2007

  79. [79]

    Computational Mechanics , volume=

    Gradient-free neural topology optimization: towards effective fracture-resistant designs , author=. Computational Mechanics , volume=. 2025 , publisher=

    work page 2025

  80. [80]

    2018 AIAA/ASCE/AHS/ASC structures, structural dynamics, and materials conference , pages=

    An indirect design representation for topology optimization using variational autoencoder and style transfer , author=. 2018 AIAA/ASCE/AHS/ASC structures, structural dynamics, and materials conference , pages=

    work page 2018

Showing first 80 references.