Pith. sign in

REVIEW 4 major objections 6 minor 81 references

EngiBench: A Framework for Data-Driven Engineering Design Research

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read EngiBench gives the field a single interface for physics-based design problems and shows that standard machine learning models fail most of them.

desk verdict A useful benchmark infrastructure with honest experiments, but the single-label-per-condition design makes the inverse-design rankings shakier than the text suggests. read the letter →

arxiv 2508.00831 v2 pith:NK4RFA7L submitted 2025-06-02 cs.CE cs.LGcs.SYeess.SY

Florian Felten , Gabriel Apaza , Gerhard Bräunlich , Cashen Diniz , Xuliang Dong , Arthur Drake , Milad Habibi , Nathaniel J. Hoffman
show 4 more authors
Matthew Keeler Soheyl Massoudi Francis G. VanGessel Mark Fuge
This is my paper
classification cs.CEcs.LGcs.SYeess.SY
keywords engineeringdesignoptimizationinversesurrogatemodelsgenerativetopologysimulation-basedbenchmarksconstraintsatisfactionreproducibility
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

EngiBench is a proposed standard platform for data-driven engineering design: one Python API wraps seven real physics-based design problems—airfoil aerodynamics, heat-conduction topology optimization in 2D and 3D, thermoelastic beams, structural beams, photonic demultiplexing, and power electronics—together with datasets, constraint checks, and optimization routines, plus a companion algorithm library. The paper's central claim is that these problems are genuinely hard for ordinary machine learning: trained generative models produce designs that mostly violate feasibility constraints or fail to simulate, and carefully tuned surrogate models produce Pareto fronts that the circuit simulator rejects. The authors attribute this to highly sensitive, constrained design manifolds where small representation changes break geometric continuity or trigger stiff, discontinuous responses. If the platform is adopted, the field gains reproducible cross-domain comparisons that were previously impractical, and the results reframe constraint satisfaction, not prediction accuracy, as the main open problem.

What carries the argument

The load-bearing mechanism is the EngiBench problem object: a versioned class that exposes design_space, objectives, conditions, dataset, check_constraints(), simulate(), optimize(), and render() through one interface. Each problem bundles a physical simulator with a precomputed dataset of designs labeled optimal_design, the conditions they were optimized under, and objective values; switching problems requires changing only the import line. The companion EngiOpt library supplies single-file implementations of generative adversarial networks, conditional GANs, diffusion models, a Bézier-parameterized GAN, and a surrogate-assisted NSGA-II stack, so the same code and metrics run across domains. This bundling lets engineering-specific metrics—cumulative optimality gap, ratio of violated constraints, ratio of failed simulations, MMD, and determinantal point process diversity—be computed uniformly, which is what makes the hardness claim measurable.

What would settle it

Take every stored optimal_design in Beams2D and HeatConduction2D, run one additional adjoint optimization step from it, and count how many improve by more than the COG gaps reported in Table 2; if many improve, the optimality metric and the hardness rankings built on it need revision.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a diverse set of realistic design problems can be brought under one interface without flattening their physics, and that standard generative and surrogate methods, once run through that interface, underperform on engineering-specific metrics. In the cross-domain inverse-design study, the unconditional GAN frequently beats a conditional GAN and a conditional diffusion model on cumulative optimality gap (COG) and maximum mean discrepancy (MMD), even though its outputs look blurrier; the diffusion model shows mode collapse on airfoils, and constraint-violation ratios on the volume-fraction constraints sit near 75 to 100 percent for most generative models. In the surrogate study, multilayer perceptron ensembles tuned by Bayesian search and searched with the standard NSGA-II multi-objective genetic algorithm produce Pareto fronts whose distributional equality with the circuit simulator's re-evaluations is rejected by MMD tests on all ten seeds, with the stiff, outlier-heavy voltage-ripple response identified as the principal cause. The conclusion the paper draws is that feasible, simulatable, optimizable output is the binding difficulty, not statistical resemblance to the training data.

Load-bearing premise

The COG optimality metric and the rankings built on it assume that the stored optimal_design entries are close enough to true optima, even though they were produced by gradient-based optimizers that can stop at local optima.

Editorial extensions

If this is right

  • Cross-domain benchmarking becomes a command-line change rather than a months-long simulator integration, so one study can compare generative and surrogate methods across half a dozen physics problems.
  • Constraint satisfaction becomes a first-class benchmark signal; models that match data distributions but violate volume fractions or crash at meshing time will be immediately visible.
  • Surrogate failures on the power-electronics circuit imply that black-box surrogates for stiff, multi-timescale systems need physics-informed features, adaptive sampling, or hybrid surrogate-simulation loops to drive optimization.
  • Airfoil results imply that representation choice—raw spline coordinates versus Bézier control points—can matter more than model architecture for producing simulatable designs.
  • Released datasets with full field data support follow-on work in physics-informed neural networks, neural operators, multi-fidelity transfer, and latent-space optimization.

Reading between the lines

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

  • A testable extension is to certify the stored designs with independent multistart optimization or perturb-and-re-optimize runs; if many labeled optima improve, the COG ranking of generative models partly reflects dataset generation artifacts rather than problem hardness.
  • The finding that blurrier unconditional GAN outputs optimize more easily suggests COG rewards outputs that are easy to refine; reporting COG jointly with constraint-satisfaction rates, or conditioning COG on feasible outputs, could flip the apparent ranking of models.
  • The multi-domain setup invites a transfer-learning probe the paper only gestures at: train a model on one problem's dataset and test on another, predicting that simulator-specific artifacts in learned representations will regress to the mean and expose overfitting.
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

4 major / 6 minor

Summary. EngiBench is presented as an open-source library and dataset collection for data-driven engineering design, providing a unified Python API, versioned problem implementations, physics simulators, constraint checking, and a companion EngiOpt library of baseline algorithms. The paper describes seven problem domains (airfoil RANS, 2D/3D heat conduction, thermo-elastic beams, beam topology optimization, photonics demultiplexing, and power electronics) and reports proof-of-concept experiments: cross-domain generative modeling (GAN, CGAN, diffusion) evaluated with COG, RVC, MMD, and DPP metrics; airfoil inverse design with simulation-failure rates; and surrogate-assisted multi-objective optimization for power electronics validated against the NgSpice simulator. The authors claim that these problems pose significant challenges for standard machine learning methods because of highly sensitive and constrained design manifolds, and they release tagged code, datasets, and experiment logs to support reproducibility.

Significance. If the benchmark's design labels and metrics are properly calibrated, EngiBench would be a valuable public resource for the engineering-design ML community: it ships tagged code (v0.0.1), HuggingFace datasets, W&B logs, Docker/Singularity support, and a uniform API that makes cross-domain comparisons straightforward. The concrete evidence for hardness includes RVC values near 1.0 for generative outputs on Beams2D and HeatConduction2D, high airfoil simulation-failure rates for a raw GAN, and surrogate Pareto fronts that are uniformly rejected by the simulator after Bayesian hyperparameter tuning and ensembling. The PowerElectronics surrogate result is particularly convincing and presented with careful validation. However, the inverse-design evaluation relies on a single stored 'optimal_design' label per condition, and the paper's own appendices document multimodality (C.5), local optima (C.4), and an unexplained dataset split (C.1); these issues must be resolved before the central claim that standard ML methods struggle due to intrinsic manifold sensitivity can be fully accepted.

major comments (4)
  1. [Section 2.1 (COG metric)] The COG metric is defined against f*, which is said to be 'typically estimated using an adjoint solver,' but the paper never states whether f* is recomputed for each test condition from a fresh optimization or read from the stored optimal_design values. If the latter, the COG results in Table 2 inherit any local-optimum bias in the dataset (see C.4), and the reported optimality gaps may not reflect intrinsic manifold difficulty. The evaluation protocol must specify how f* is obtained, including the optimizer settings, starting points, and condition set, and should preferably recompute f* independently of the dataset labels.
  2. [Section C.5 / Table 2 (Photonics2D)] Section C.5 explicitly states that for Photonics2D 'there are usually multiple solutions with equivalent or similar performance' and 'the solution may not have a single unique global minimum,' yet the dataset stores exactly one optimal_design per condition. Under this labeling, conditional generative models are penalized by MMD/DPP and often by RVC/RF for generating a different valid optimum, so the high failure rates and poor COG reported for P2D in Table 2 may be artifacts of single-mode labeling rather than measures of design-manifold sensitivity. The authors should provide multiple optima per condition, quantify the spread of equivalent optima, or reformulate the inverse-design evaluation to be invariant under equivalent-solution sets before the photonics hardness result is used as evidence for the central claim.
  3. [Section C.4 / Table 2 (Beams2D)] Appendix C.4 reports that the OC inner-loop termination condition 'prevents the code from becoming stuck at this point, which we observed in some warm-starting instances with noisy initial designs,' indicating that some dataset designs may be local optima rather than true optima. Because the Beams2D COG and RVC metrics in Table 2 depend on the stored optimal_design labels (and possibly on f* derived from them), the paper should quantify how many dataset entries were affected by such warm-start failures, state whether these entries were retained, and verify that excluding them does not change the qualitative conclusions.
  4. [Section C.1 (Airfoil dataset)] Section C.1 reports 1400 parameter combinations sampled by Latin hypercube sampling but then describes a training/validation/test split of 748+140+47=935 samples, leaving 465 samples unaccounted for. Because the Airfoil experiments in Section 4.2 (RF values for GAN, diffusion, and BézierGAN) use this dataset, the authors must explain the missing samples, document any filtering or removal criteria, and confirm that the split was random and not selective. Unreported filtering could bias the reported simulation-failure rates and thus the conclusion that domain-informed models improve performance.
minor comments (6)
  1. [Section C.5] The Dataset paragraph contains a typo: 'sampling by sampling at random' should read 'sampling at random'.
  2. [Abstract] The phrase 'the first open-source library and datasets spanning diverse domains' should be qualified relative to existing multi-domain simulation benchmarks such as The Well and PDEBench, which also provide public data and APIs even if they are not focused on engineering design optimization; consider softening 'first' or explicitly distinguishing the design-optimization focus.
  3. [References [72]-[73]] Reference [73] (Neil Wu et al.) duplicates the title of reference [72] (Ella Wu et al.); both appear to be for pyOptSparse, and the intended citation should be checked and disambiguated.
  4. [Table 2 caption] The caption lists four metrics per cell but the P2D rows show 'N/A' for RVC without explanation; a footnote stating that Photonics2D has no volume-fraction constraint would help readers avoid confusion.
  5. [Section 4.1 / Appendix D.1] The authors note in Appendix D.1 that they did not perform extensive hyperparameter tuning and that ten seeds are insufficient for strong statistical claims; this caveat should also appear in the main text near Table 2, since otherwise the table may be read as a definitive ranking of generative models rather than a proof-of-concept demonstration.
  6. [Appendix D.3] The MMD permutation test is described as using 1000 permutations with p-values hitting a floor of 0.001; the authors should clarify whether the null distribution was computed with a pooled-sample permutation procedure or a bootstrap so that the test is exactly reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: EngiBench's benchmark claims are empirical and independently grounded in external simulators; self-citations are provenance, not load-bearing reductions.

full rationale

EngiBench is not a derivational paper: its central claims are that the unified API makes cross-domain benchmarking feasible and that standard ML methods struggle on these constrained design problems. Those claims are supported by direct simulation with external, independently developed solvers (MachAero/ADflow for Airfoil, Dolfin-adjoint/Ipopt for heat conduction, ceviche for Photonics2D, NgSpice for PowerElectronics) and by standard metrics (MMD, DPP, RVC, RF, COG) that are not defined in terms of the models being evaluated. The COG baseline f* is described in Section 2.1 as 'the optimal objective value under condition c, typically estimated using an adjoint solver,' which is an external reference rather than a fitted parameter; nothing in the paper defines f* as the stored dataset label in a way that would make the optimality comparison tautological. The paper's self-citations—for example Diniz and Fuge [13] for the Airfoil dataset, Habibi et al. [24,25] for the heat-conduction datasets, and Chen et al. [11] for BézierGAN—are provenance for data and algorithms, not an argument whose conclusion is assumed. The admitted limitations, including Photonics2D's multiple equivalent optima (Section C.5) and Beams2D's optimizer warm-start convergence issues (Section C.4), are benchmark-validity concerns about label well-posedness, not circular derivations; likewise the Airfoil sample-count discrepancy in Section C.1 is a reporting issue. No equation or metric in the paper reduces by construction to its own input, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on the construction of the benchmarks and the reliability of the generated datasets. No free parameters are fit to claim a new physical law, but several hand-set constants define problem difficulty. No invented scientific entities are introduced. The main assumptions are simulator fidelity, near-optimality of stored designs, split integrity, and kernel choices for the metrics.

free parameters (4)
  • Photonics2D material usage penalty w = 1e-2
    Hand-set default in the objective (Section C.5); directly changes which designs are optimal and therefore the difficulty of inverse design.
  • Photonics2D beta continuation schedule = beta from 1.0 to 300.0, quadratic
    Hand-chosen continuation that binarizes the material distribution (Section C.5); influences the quality and topology of dataset designs.
  • HeatConduction2D/3D condition sampling bounds = volume in [0.3, 0.6]; adiabatic length in [0,1]; area in [0,1.0]
    Appendix C.2 states these ranges; they define the benchmark's feasible conditions and were not derived from a physical law.
  • Baseline hyperparameters (Tables 8-11) = e.g., lr_disc 4e-4, lr_gen 1e-4, n_epochs 100; diffusion lr 3e-4; seeds 1..10
    Chosen from prior experience without extensive tuning (Appendix D.1); the empirical ordering of models in Section 4.1 is sensitive to these settings and the authors caution against overinterpretation.
assumptions (5)
  • domain assumption The physics simulators (ADflow/RANS, FEniCS, ceviche, NgSpice) produce objective labels accurate enough to rank designs
    Section 3.3 and Appendix C define every problem through these simulators; if simulation errors dominate, the benchmark's ground-truth labels and all derived metrics are unreliable.
  • domain assumption Dataset 'optimal_design' entries are close enough to the true constrained optima to support COG comparisons
    Section 2.1 defines f* as 'typically estimated using an adjoint solver'; the datasets were generated with gradient-based methods (SLSQP, MMA, OC, Adam, Ipopt) that can converge to local optima, and the Airfoil sample count inconsistency (1400 vs 935) raises further questions about how designs were kept.
  • domain assumption Train/test splits of the datasets do not leak information and are representative of the design-condition space
    Section C.2 explicitly excludes some parameter values from training for HeatConduction; Section C.1 describes only a random split for Airfoil and does not state whether conditions overlap between splits; any leakage would inflate or deflate measured difficulty.
  • domain assumption The MMD and DPP kernels (Gaussian) and their bandwidths are applied consistently and meaningfully across problems
    Section 2.1 defines MMD with an unspecified Gaussian kernel and DPP with a similarity kernel; Appendix D uses 'a Gaussian kernel' without specifying bandwidth, so metric values may not be reproducible or comparable across problems.
  • ad hoc to paper The hand-set ranges for conditions (e.g., volume fraction bounds, rmin, beta continuation) define a fair and representative difficulty level
    Appendix C sets bounds such as volume in [0.3, 0.6] for HeatConduction2D and beta in [1,300] for Photonics2D; these choices are not derived from a physical requirement and shape the benchmark's difficulty.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EngiBench: A Framework for Data-Driven Engineering Design Research." pith.science (2026). https://pith.science/paper/NK4RFA7L

@misc{pith2026250800831,
  author       = {Pith},
  title        = {Pith review of: EngiBench: A Framework for Data-Driven Engineering Design Research},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NK4RFA7L}},
  note         = {Machine review of arXiv:2508.00831}
}
read the original abstract

Engineering design optimization seeks to automatically determine the shapes, topologies, or parameters of components that maximize performance under given conditions. This process often depends on physics-based simulations, which are difficult to install, computationally expensive, and require domain-specific expertise. To mitigate these challenges, we introduce EngiBench, the first open-source library and datasets spanning diverse domains for data-driven engineering design. EngiBench provides a unified API and a curated set of benchmarks -- covering aeronautics, heat conduction, photonics, and more -- that enable fair, reproducible comparisons of optimization and machine learning algorithms, such as generative or surrogate models. We also release EngiOpt, a companion library offering a collection of such algorithms compatible with the EngiBench interface. Both libraries are modular, letting users plug in novel algorithms or problems, automate end-to-end experiment workflows, and leverage built-in utilities for visualization, dataset generation, feasibility checks, and performance analysis. We demonstrate their versatility through experiments comparing state-of-the-art techniques across multiple engineering design problems, an undertaking that was previously prohibitively time-consuming to perform. Finally, we show that these problems pose significant challenges for standard machine learning methods due to highly sensitive and constrained design manifolds.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

81 extracted references · 54 canonical work pages

  1. [1]

    http://ngspice.sourceforge.net, 2025

    Ngspice: A mixed-level/mixed-signal circuit simulator. http://ngspice.sourceforge.net, 2025. Accessed: 2025-03-31. 4, 8, 31

    2025

  2. [2]

    Lazarov, and Ole Sigmund

    Erik Andreassen, Anders Clausen, Mattias Schevenels, Boyan S. Lazarov, and Ole Sigmund. Efficient topology optimization in MATLAB using 88 lines of code. Structural and Multidisciplinary Optimization, 43(1):1–16, January 2011. ISSN 1615-1488. doi: 10.1007/s00158-010-0594-7. URL https://doi. org/10.1007/s00158-010-0594-7 . 6, 7, 26, 27, 28

    doi:10.1007/s00158-010-0594-7 2011

  3. [3]

    Botorch: A framework for efficient monte-carlo bayesian optimization

    Maximilian Balandat, Brian Karrer, Daniel Jiang, Samuel Daulton, Ben Letham, Andrew G Wilson, and Eytan Bakshy. Botorch: A framework for efficient monte-carlo bayesian optimization. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, 2020. 6, 10, 41

    work page 2020

  4. [4]

    Gantl: Toward practical and real-time topology optimization with conditional generative adversarial networks and transfer learning

    Mohammad Mahdi Behzadi and Horea T Ilie¸ s. Gantl: Toward practical and real-time topology optimization with conditional generative adversarial networks and transfer learning. Journal of Mechanical Design, 144 (2):021711, 2022. 2

    work page 2022

  5. [5]

    Optimal shape design as a material distribution problem

    Martin P Bendsøe. Optimal shape design as a material distribution problem. Structural optimization, 1(4): 193–202, 1989. 27

    work page 1989

  6. [6]

    Topology optimization: theory, methods, and applications

    Martin Philip Bendsoe and Ole Sigmund. Topology optimization: theory, methods, and applications . Springer Science & Business Media, 2013. 23

    work page 2013

  7. [7]

    Experiment tracking with weights and biases, 2020

    Lukas Biewald. Experiment tracking with weights and biases, 2020. URL https://www.wandb.com/. Software available from wandb.com. 10

    work page 2020

  8. [8]

    Blank and K

    J. Blank and K. Deb. pymoo: Multi-Objective Optimization in Python. IEEE Access, 8:89497–89509,

    work page

Show all 81 references
  1. [9]

    Inverse design of two-dimensional airfoils using conditional generative models and surrogate log-likelihoods

    Qiuyi Chen, Jun Wang, Phillip Pope, Wei Chen, and Mark Fuge. Inverse design of two-dimensional airfoils using conditional generative models and surrogate log-likelihoods. Journal of Mechanical Design, 144(2): 021712, 2022. 2

    2022

  2. [10]

    Aerodynamic Design Optimization and Shape Exploration using Generative Adversarial Networks

    Wei Chen, Kevin Chiu, and Mark Fuge. Aerodynamic Design Optimization and Shape Exploration using Generative Adversarial Networks. 2019. doi: 10.2514/6.2019-2351. URL https://arc.aiaa.org/ doi/abs/10.2514/6.2019-2351. 4, 18

    2019 doi

  3. [11]

    Wei Chen, Kevin Chiu, and Mark D. Fuge. Airfoil Design Parameterization and Optimization Using Bézier Generative Adversarial Networks. AIAA Journal, 58(11):4723–4735, 2020. ISSN 0001-1452. doi: 10.2514/1.J059317. URL https://doi.org/10.2514/1.J059317. Publisher: American Insti...

    2020 doi

  4. [12]

    K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation, 6(2):182–197, April 2002. ISSN 1941-0026. doi: 10.1109/4235.996017. 8, 41

    2002

  5. [13]

    Optimizing diffusion to diffuse optimal designs

    Cashen Diniz and Mark Fuge. Optimizing diffusion to diffuse optimal designs. In AIAA SCITECH 2024 Forum. American Institute of Aeronautics and Astronautics, 2024. doi: 10.2514/6.2024-2013. URL https://arc.aiaa.org/doi/abs/10.2514/6.2024-2013. 6, 7, 18, 21

    2024 doi

  6. [14]

    Hyperparameters in Reinforcement Learning and How To Tune Them

    Theresa Eimer, Marius Lindauer, and Roberta Raileanu. Hyperparameters in Reinforcement Learning and How To Tune Them. In Proceedings of the 40th International Conference on Machine Learning (ICML 2023), June 2023. URL https://openreview.net/forum?id=0Vm8Ghcxmp. 41

    2023

  7. [15]

    Drivaernet++: A large-scale multimodal car dataset with computational fluid dynamics simulations and deep learning benchmarks

    Mohamed Elrefaie, Florin Morar, Angela Dai, and Faez Ahmed. Drivaernet++: A large-scale multimodal car dataset with computational fluid dynamics simulations and deep learning benchmarks. In Advances in Neural Information Processing Systems, volume 37, pages 499–536, 2024. 2, 4

    2024

  8. [16]

    Topology optimization of heat exchangers: A review

    Ahmad Fawaz, Yuchao Hua, Steven Le Corre, Yilin Fan, and Lingai Luo. Topology optimization of heat exchangers: A review. Energy, 252:124053, 2022. 22

    2022

  9. [17]

    Florian Felten, Lucas Nunes Alegre, Ann Nowe, Ana L. C. Bazzan, El Ghazali Talbi, Grégoire Danoy, and Bruno Castro da Silva. A Toolkit for Reliable Benchmarking and Research in Multi-Objective Reinforcement Learning. In Proceedings of the 37th Conference on Neural Information ...

    2023

  10. [18]

    Hyperparameter Optimization for Multi-Objective Reinforcement Learning, October 2023

    Florian Felten, Daniel Gareev, El-Ghazali Talbi, and Grégoire Danoy. Hyperparameter Optimization for Multi-Objective Reinforcement Learning, October 2023. URL http://arxiv.org/abs/2310.16487. arXiv:2310.16487 [cs]. 41

    2023 arXiv

  11. [19]

    Roijers, Jordan K

    Florian Felten, Umut Ucak, Hicham Azmani, Gao Peng, Willem Röpke, Hendrik Baier, Patrick Mannion, Diederik M. Roijers, Jordan K. Terry, El-Ghazali Talbi, Grégoire Danoy, Ann Nowé, and Roxana R˘adulescu. MOMAland: A Set of Benchmarks for Multi-Objective Multi-Agent Reinforcemen...

    2024 arXiv

  12. [20]

    M. A. Ganaie, Minghui Hu, A. K. Malik, M. Tanveer, and P. N. Suganthan. Ensemble deep learning: A review. Engineering Applications of Artificial Intelligence, 115:105151, October 2022. ISSN 0952-1976. doi: 10.1016/j.engappai.2022.105151. 9, 41

    2022

  13. [21]

    Multi-material thermo- mechanical topology optimization with applications to additive manufacturing: Design of main composite part and its support structure

    Oliver Giraldo-Londoño, Lucia Mirabella, Livio Dalloro, and Glaucio H Paulino. Multi-material thermo- mechanical topology optimization with applications to additive manufacturing: Design of main composite part and its support structure. Computer Methods in Applied Mechanics an...

  14. [22]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. URL https://proceedings.neurips.cc/paper_files/paper/20...

    2014

  15. [23]

    A kernel two-sample test

    Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test. The Journal of Machine Learning Research, 13(1):723–773, 2012. 3, 41

    2012

  16. [24]

    Milad Habibi, Jun Wang, and Mark Fuge. When is it actually worth learning inverse design? In International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, volume 87301, page V03AT03A025. American Society of Mechanical Engineers...

    2023

  17. [25]

    Mean squared error may lead you astray when optimizing your inverse design methods

    Milad Habibi, Shai Bernard, Jun Wang, and Mark Fuge. Mean squared error may lead you astray when optimizing your inverse design methods. Journal of Mechanical Design, 147(2):021701, 2025. 3, 6, 23

    2025

  18. [26]

    Hajdik, Anil Yildirim, Neil Wu, Benjamin J

    Hannah M. Hajdik, Anil Yildirim, Neil Wu, Benjamin J. Brelje, Sabet Seraj, Marco Mangano, Joshua L. Anibal, Eirikur Jonsson, Eytan J. Adler, Charles A. Mader, Gaetan K. W. Kenway, and Joaquim R. R. A. Martins. pyGeo: A geometry package for multidisciplinary design optimization...

    2023 doi

  19. [27]

    Harris, K

    Charles R. Harris, K. Jarrod Millman, Stéfan J. van der Walt, Ralf Gommers, Pauli Virtanen, David Cournapeau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J. Smith, Robert Kern, Matti Picus, Stephan Hoyer, Marten H. van Kerkwijk, Matthew Brett, Allan Haldane, Jaime Fe...

    2020

  20. [28]

    Bubbleml: A multi-physics dataset and benchmarks for machine learning

    Sheikh Md Shakeel Hassan, Arthur Feeney, Akash Dhruv, Jihoon Kim, Youngjoon Suh, Jaiyoung Ryu, Yoonjin Won, and Aparna Chandramowlishwaran. Bubbleml: A multi-physics dataset and benchmarks for machine learning. arXiv preprint arXiv:2307.14623, 2023. 4

    2023 arXiv

  21. [29]

    Mader, Anil Yildirim, and Joaquim R

    Xiaolong He, Jichao Li, Charles A. Mader, Anil Yildirim, and Joaquim R. R. A. Martins. Robust aerodynamic shape optimization—from a circle to an airfoil. Aerospace Science and Technology, 87:48–61, April 2019. doi: 10.1016/j.ast.2019.01.051. 18

    2019 doi

  22. [30]

    Denoising Diffusion Probabilistic Models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In Advances in Neural Information Processing Systems , volume 33, pages 6840–6851. Cur- ran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper/2020/hash/ 4c5bcfec8584af0d967f1ab1...

    2020

  23. [31]

    Shengyi Huang, Rousslan Fernand Julien Dossa, Chang Ye, Jeff Braga, Dipam Chakraborty, Kinal Mehta, and João G. M. Araújo. CleanRL: High-quality Single-file Implementations of Deep Reinforcement Learning Algorithms. Journal of Machine Learning Research, 23(274):1–18, 2022. ISS...

    2022

  24. [32]

    Adjoint method and inverse design for nonlinear nanophotonic devices

    Tyler W Hughes, Momchil Minkov, Ian AD Williamson, and Shanhui Fan. Adjoint method and inverse design for nonlinear nanophotonic devices. ACS Photonics, 5(12):4781–4787, 2018. 7, 29, 30

    2018

  25. [33]

    Forward-mode differentiation of maxwell’s equations

    Tyler W Hughes, Ian AD Williamson, Momchil Minkov, and Shanhui Fan. Forward-mode differentiation of maxwell’s equations. ACS Photonics, 6(11):3010–3016, 2019. 7, 30

    2019

  26. [34]

    Gaetan K. W. Kenway, Charles A. Mader, Ping He, and Joaquim R. R. A. Martins. Effective adjoint approaches for computational fluid dynamics. Progress in Aerospace Sciences, 110:100542, October 2019. doi: 10.1016/j.paerosci.2019.05.002. 20

    2019 doi

  27. [35]

    Shiphullgan: A generic parametric modeller for ship hull design using deep convolutional generative model

    Shahroz Khan, Kosa Goucher-Lambert, Konstantinos Kostas, and Panagiotis Kaklis. Shiphullgan: A generic parametric modeller for ship hull design using deep convolutional generative model. Computer Methods in Applied Mechanics and Engineering, 411:116051, 2023. 2

    2023

  28. [36]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 4

    2009

  29. [37]

    Determinantal point processes for machine learning

    Alex Kulesza and Ben Taskar. Determinantal point processes for machine learning. Foundations and Trends® in Machine Learning, 5(2–3):123–286, 2012. ISSN 1935-8237. doi: 10.1561/2200000044. URL http://dx.doi.org/10.1561/2200000044. 4 12

    2012 doi

  30. [38]

    Kurtzer, cclerget, Michael Bauer, Ian Kaneshiro, David Trudgian, and David Godlove

    Gregory M. Kurtzer, cclerget, Michael Bauer, Ian Kaneshiro, David Trudgian, and David Godlove. hpcng/singularity: Singularity 3.7.3, April 2021. URL https://doi.org/10.5281/zenodo.4667718. 10

    2021 doi

  31. [39]

    Mnist handwritten digit database

    Yann LeCun, Corinna Cortes, and CJ Burges. Mnist handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, 2, 2010. 4

    2010

  32. [40]

    Datasets: A Community Library for Natural Language Processing

    Quentin Lhoest, Albert Villanova del Moral, Patrick von Platen, Thomas Wolf, Mario Šaško, Yacine Jernite, Abhishek Thakur, Lewis Tunstall, Suraj Patil, Mariama Drame, Julien Chaumond, Julien Plu, Joe Davison, Simon Brandeis, Victor Sanh, Teven Le Scao, Kevin Canwen Xu, Nicolas...

    2021

  33. [41]

    Anders Logg and Garth N. Wells. Dolfin: Automated finite element computing. ACM Trans. Math. Softw., 37(2), April 2010. ISSN 0098-3500. doi: 10.1145/1731022.1731030. URL https://doi.org/10. 1145/1731022.1731030. 4, 7, 10

    2010

  34. [42]

    Automated Solution of Differential Equations by the Finite Element Method: The FEniCS Book, volume 84 of Lecture Notes in Computational Science and Engineering

    Anders Logg, Kent-Andre Mardal, and Garth Wells, editors. Automated Solution of Differential Equations by the Finite Element Method: The FEniCS Book, volume 84 of Lecture Notes in Computational Science and Engineering. Springer, Berlin, Heidelberg, 2012. ISBN 978-3-642-23098-1...

    2012 doi

  35. [43]

    Mader, Gaetan K

    Charles A. Mader, Gaetan K. W. Kenway, Anil Yildirim, and Joaquim R. R. A. Martins. ADflow—an open-source computational fluid dynamics solver for aerodynamic and multidisciplinary optimization. Journal of Aerospace Information Systems, 2020. doi: 10.2514/1.I010796. 4, 7, 10, 20

    2020 doi

  36. [44]

    Aerodynamic design optimization: Challenges and perspectives

    Joaquim Martins. Aerodynamic design optimization: Challenges and perspectives. Computers & Fluids, 239:105391, 03 2022. doi: 10.1016/j.compfluid.2022.105391. 18

    2022

  37. [45]

    Joaquim R. R. A. Martins and Andrew Ning. Engineering Design Optimization. Cambridge University Press, Cambridge, UK, January 2022. ISBN 9781108833417. doi: 10.1017/9781108980647. URL https://mdobook.github.io. 6, 18

    2022 doi

  38. [46]

    Dimensionless group-driven ensemble neural networks for robust design optimization in engineering

    Soheyl Massoudi and Jürg Schiffmann. Dimensionless group-driven ensemble neural networks for robust design optimization in engineering. Journal of Computational Design and Engineering, 12(7):61–95, July

  39. [47]

    An Integrated Approach to Designing Robust Gas-Bearing Supported Turbocompressors Through Surrogate Modeling and Constrained All-At-Once Multi-Objective Optimization

    Soheyl Massoudi, Cyril Picard, and Jürg Schiffmann. An Integrated Approach to Designing Robust Gas-Bearing Supported Turbocompressors Through Surrogate Modeling and Constrained All-At-Once Multi-Objective Optimization. Journal of Mechanical Design, 146(121706), July 2024. ISSN...

    2024 doi

  40. [48]

    Robust design optimization of gas-lubricated herringbone grooved journal bearings: Surrogate modeling and experimental validation

    Soheyl Massoudi, Cameron Bush, and Jürg Schiffmann. Robust design optimization of gas-lubricated herringbone grooved journal bearings: Surrogate modeling and experimental validation. Tribology International, 204:110429, April 2025. ISSN 0301-679X. doi: 10.1016/j.triboint.2024....

    2025

  41. [49]

    From topology optimization design to additive manufacturing: Today’s success and tomorrow’s roadmap

    Liang Meng, Weihong Zhang, Dongliang Quan, Guanghui Shi, Lei Tang, Yuliang Hou, Piotr Breitkopf, Jihong Zhu, and Tong Gao. From topology optimization design to additive manufacturing: Today’s success and tomorrow’s roadmap. Archives of Computational Methods in Engineering, 27(...

    2020

  42. [50]

    Docker: lightweight linux containers for consistent development and deployment

    Dirk Merkel. Docker: lightweight linux containers for consistent development and deployment. Linux J., 2014(239), March 2014. ISSN 1075-3583. 10

    2014

  43. [51]

    Conditional Generative Adversarial Nets, November 2014

    Mehdi Mirza and Simon Osindero. Conditional Generative Adversarial Nets, November 2014. URL http://arxiv.org/abs/1411.1784. arXiv:1411.1784 [cs]. 8

    2014 arXiv

  44. [52]

    dolfin-adjoint 2018.1: automated adjoints for fenics and firedrake

    Sebastian Mitusch, Simon Funke, and Jørgen Dokken. dolfin-adjoint 2018.1: automated adjoints for fenics and firedrake. Journal of Open Source Software, 4(38):1292, 2019. 7, 10, 23

    2018

  45. [53]

    Topology optimization of cooling plates for battery thermal management

    Xiaobao Mo, Hui Zhi, Yizhi Xiao, Haiyu Hua, and Liang He. Topology optimization of cooling plates for battery thermal management. International Journal of Heat and Mass Transfer, 178:121612, 2021. 22

    2021

  46. [54]

    The well: a large-scale collection of diverse physics simulations for machine learning

    Ruben Ohana, Michael McCabe, Lucas Meyer, Rudy Morel, Fruzsina Agocs, Miguel Beneitez, Marsha Berger, Blakesly Burkhart, Stuart Dalziel, Drummond Fielding, et al. The well: a large-scale collection of diverse physics simulations for machine learning. Advances in Neural Informa...

    2024

  47. [55]

    Automated Reinforcement Learning (AutoRL): A Survey and Open Problems

    Jack Parker-Holder, Raghu Rajan, Xingyou Song, André Biedenkapp, Yingjie Miao, Theresa Eimer, Baohe Zhang, Vu Nguyen, Roberto Calandra, Aleksandra Faust, Frank Hutter, and Marius Lindauer. Automated Reinforcement Learning (AutoRL): A Survey and Open Problems. Journal of Artifi...

    2022 doi

  48. [56]

    PyTorch: An Imperative Style, High- Performance Deep Learning Library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu ...

    2019

  49. [57]

    Inverse design and demonstration of a compact and broadband on-chip wavelength demultiplexer

    Alexander Y Piggott, Jesse Lu, Konstantinos G Lagoudakis, Jan Petykiewicz, Thomas M Babinec, and Jelena Vuˇckovi´c. Inverse design and demonstration of a compact and broadband on-chip wavelength demultiplexer. Nature photonics, 9(6):374–377, 2015. 7, 29, 30

    2015

  50. [58]

    Unsupervised representation learning with deep convo- lutional generative adversarial networks

    Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convo- lutional generative adversarial networks. In Yoshua Bengio and Yann LeCun, editors,4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, M...

    2016 arXiv

  51. [59]

    BIKED: A Dataset for Computational Bicycle Design With Machine Learning Benchmarks

    Lyle Regenwetter, Brent Curry, and Faez Ahmed. BIKED: A Dataset for Computational Bicycle Design With Machine Learning Benchmarks. Journal of Mechanical Design, 144(031706), October 2021. ISSN 1050-0472. doi: 10.1115/1.4052585. URL https://doi.org/10.1115/1.4052585. 4

    2021 doi

  52. [60]

    Deep generative models in engineering design: A review

    Lyle Regenwetter, Amin Heyrani Nobari, and Faez Ahmed. Deep generative models in engineering design: A review. Journal of Mechanical Design, 144(7), 2022. 2, 3

    2022

  53. [61]

    Beyond Statistical Similarity: Rethinking Metrics for Deep Generative Models in Engineering Design

    Lyle Regenwetter, Akash Srivastava, Dan Gutfreund, and Faez Ahmed. Beyond Statistical Similarity: Rethinking Metrics for Deep Generative Models in Engineering Design. Computer-Aided Design, 165: 103609, December 2023. ISSN 0010-4485. doi: 10.1016/j.cad.2023.103609. URL https:/...

    2023

  54. [62]

    Ney Secco, Gaetan K. W. Kenway, Ping He, Charles A. Mader, and Joaquim R. R. A. Martins. Efficient mesh generation and deformation for aerodynamic shape optimization. AIAA Journal, 2021. doi: 10.2514/ 1.J059491. 4, 7, 10, 20

    2021

  55. [63]

    Morphology-based black and white filters for topology optimization

    Ole Sigmund. Morphology-based black and white filters for topology optimization. Structural and Multidisciplinary Optimization, 33(4):401–424, 2007. 25, 27

    2007

  56. [64]

    Practical Bayesian Optimization of Machine Learning Algorithms

    Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical Bayesian Optimization of Machine Learning Algorithms. In Advances in Neural Information Processing Systems, volume 25. Curran Associates, Inc.,

  57. [65]

    Neural networks for topology optimization

    Ivan Sosnovik and Ivan Oseledets. Neural networks for topology optimization. Russian Journal of Numerical Analysis and Mathematical Modelling , 34(4):215–223, August 2019. ISSN 1569-

    2019

  58. [66]

    Pdebench: An extensive benchmark for scientific machine learning.Advances in Neural Information Processing Systems, 35:1596–1611, 2022

    Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Daniel MacKinlay, Francesco Alesiani, Dirk Pflüger, and Mathias Niepert. Pdebench: An extensive benchmark for scientific machine learning.Advances in Neural Information Processing Systems, 35:1596–1611, 2022. 4

    2022

  59. [67]

    Topology optimization of nonlinear heat conduction problems involving large temperature gradient

    Lei Tang, Tong Gao, Longlong Song, Liang Meng, Chengqi Zhang, and Weihong Zhang. Topology optimization of nonlinear heat conduction problems involving large temperature gradient. Computer Methods in Applied Mechanics and Engineering, 357:112600, 2019. 22

    2019

  60. [68]

    Lagrangebench: A lagrangian fluid mechanics benchmarking suite

    Artur Toshev, Gianluca Galletti, Fabian Fritz, Stefan Adami, and Nikolaus Adams. Lagrangebench: A lagrangian fluid mechanics benchmarking suite. Advances in Neural Information Processing Systems, 36: 64857–64884, 2023. 4

    2023

  61. [69]

    Mark Towers, Ariel Kwiatkowski, Jordan Terry, John U. Balis, Gianluca De Cola, Tristan Deleu, Manuel Goulão, Andreas Kallinteris, Markus Krimmel, Arjun KG, Rodrigo Perez-Vicente, Andrea Pierré, Sander Schulhoff, Jun Jet Tai, Hannah Tan, and Omar G. Younis. Gymnasium: A Standar...

    2024 arXiv

  62. [70]

    Diffusers: State-of-the-art diffusion models

    Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, Steven Liu, William Berman, Yiyi Xu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. URL https://github.com/huggingface/diffuser...

  63. [71]

    On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming

    Andreas Wächter and Lorenz T Biegler. On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming. Mathematical programming, 106:25–57, 2006. 23 14

    2006

  64. [73]

    Mader, John Jasa, and Joaquim R

    Neil Wu, Gaetan Kenway, Charles A. Mader, John Jasa, and Joaquim R. R. A. Martins. pyoptsparse: A python framework for large-scale constrained nonlinear optimization of sparse systems. Journal of Open Source Software, 5(54):2564, 2020. doi: 10.21105/joss.02564. 20

    2020 doi

  65. [74]

    Deep ensemble neural-like P systems for segmentation of central serous chorioretinopathy lesion

    Jie Xue, Zhuo Wang, Deting Kong, Yuan Wang, Xiyu Liu, Wen Fan, Songtao Yuan, Sijie Niu, and Dengwang Li. Deep ensemble neural-like P systems for segmentation of central serous chorioretinopathy lesion. Information Fusion, 65:84–94, January 2021. ISSN 1566-2535. doi: 10.1016/j....

    2021 doi

  66. [75]

    A jacobian-free approximate newton–krylov startup strategy for rans simulations

    Anil Yildirim, Gaetan KW Kenway, Charles A Mader, and Joaquim RRA Martins. A jacobian-free approximate newton–krylov startup strategy for rans simulations. Journal of Computational Physics, 397: 108741, 2019. 20

    2019

  67. [76]

    compliance

    Andy B. Yoo, Morris A. Jette, and Mark Grondona. SLURM: Simple Linux Utility for Resource Manage- ment. In Dror Feitelson, Larry Rudolph, and Uwe Schwiegelshohn, editors, Job Scheduling Strategies for Parallel Processing, Lecture Notes in Computer Science, pages 44–60, Berlin,...

    2003 doi

  68. [79]

    Voltage Rippleis intrinsically harder to approximate than DcGain due to stiffness (high fre- quency transients arise from stiff systems dynamics, a regime where standard MLPs struggle to generalize accurately), discontinuities, and skewed distributions driven by switching and ...

  69. [80]

    Ensembling reduces variance but does not resolve extrapolation bias in undersampled or near-discontinuous regions

  70. [81]

    Sparse representation of unstable behaviors ( e.g., large ripple) and overfitting in smooth regions can mislead optimization, resulting in systematic surrogate error near the predicted Pareto front

  71. [82]

    Domain-aware enhancements—such as physics-informed architectures, adaptive sampling near switching thresholds, or hybrid surrogate-simulation loops—may be needed to reliably optimize high-fidelity circuit models. Future work could explore hybrid strategies that combine adaptiv...

  72. [2025]

    doi: 10.1093/jcde/qwaf056

    ISSN 2288-5048. doi: 10.1093/jcde/qwaf056. 8

    doi

  73. [3988]

    URL https://www.degruyter.com/document/doi/10.1515/ rnam-2019-0018/html

    doi: 10.1515/rnam-2019-0018. URL https://www.degruyter.com/document/doi/10.1515/ rnam-2019-0018/html . Publisher: De Gruyter. 4

    2019 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.