Pith. sign in

REVIEW 3 major objections 5 minor 69 references

Physics-guided diffusion cuts topology optimization errors to 5% on unseen loads

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 16:30 UTC pith:PSXYRGKT

load-bearing objection Solid generative TO method with two real contributions; headline OOD number is best-of-5-seeds, inflating margins by ~21% the 3 major comments →

arxiv 2607.07233 v1 pith:PSXYRGKT submitted 2026-07-08 cs.LG cs.CE

HPG-Diff: Hierarchical physics-guided diffusion with differentiable connectivity constraints for topology optimization

Jinbo Yang , Mingyue Yuan , Boyuan Zhang , Yoshifumi Kitamura , Shikai Jing This is my paper
classification cs.LG cs.CE
keywords topology optimizationdiffusion modelsphysics-guided generationconnectivity constraintsstructural mechanicsgenerative designdifferentiable constraints
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 proposes HPG-Diff, a diffusion model for structural topology optimization that matches the physics fidelity of classical iterative solvers while retaining the speed and diversity of generative models. The central mechanism is a hierarchical injection of precomputed mechanical features—displacement fields, principal stress lines, and strain energy density—into different depth levels of the denoising UNet, so that each layer receives the physics signal whose spatial scale matches that layer's receptive field. A second mechanism, the Floating Material Suppression (FMS) loss, reformulates the discrete connectivity problem as a differentiable heat-propagation process: a virtual heat source is placed at the load position, and any high-density material that remains cold after propagation is penalized as structurally disconnected. Together these two mechanisms reduce compliance errors to under 1% on familiar boundary conditions and 5.3% on unseen ones, while cutting the rate of disconnected floating material to roughly 2.4%, all without external guidance models or post-processing filters.

Core claim

The paper claims that explicitly organizing physics features by their spatial abstraction level—low-level displacement for shallow layers, mid-level stress lines for middle layers, high-level strain energy for deep layers—substantially improves a diffusion model's ability to generalize to unseen structural design problems. This hierarchical alignment, combined with a differentiable thermal-conduction penalty for disconnected material, allows a generative model to approach the mechanical fidelity of classical finite-element-based optimization at a fraction of the cost, producing multiple valid design alternatives in seconds rather than minutes.

What carries the argument

The framework rests on two coupled mechanisms. First, cross-attention modules at designated UNet layers inject the corresponding physics feature: displacement (U) at shallow layers for fine-grained local detail, principal stress lines (PSL) at mid layers for load-path skeleton, and strain energy density (SED) at deep layers for global stiffness-critical regions. Second, the FMS loss simulates iterative 3x3 max-dilation propagation from a load-seed mask through the predicted density field; regions with high predicted density but low propagated connectivity signal receive gradient penalties. An exponential time-weighting confines the FMS penalty to late denoising steps, preventing gradient冲突s.

Load-bearing premise

The paper assumes that the specific assignment of displacement to shallow layers, stress lines to mid layers, and strain energy to deep layers is the correct and robust alignment. The evidence for this is largely ablation-based on a single UNet architecture, and the orthogonality analysis reveals that two of the three features (displacement and stress lines) are highly correlated, leaving open whether the claimed layer-feature matching is a general principle or an artifact of

What would settle it

If the hierarchical feature-to-layer assignment provides no benefit beyond simply concatenating all three features and injecting them uniformly—as tested in the 'Bottleneck' ablation—the core architectural claim collapses. The paper reports this ablation does show degradation, but the margin on in-distribution data is modest, and the out-of-distribution improvement could partly stem from the increased conditioning capacity rather than the hierarchical structure itself.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Generative topology optimization could replace classical solvers in early-stage design exploration where multiple structurally valid alternatives are needed quickly, not just a single optimum.
  • The differentiable heat-propagation connectivity penalty could be transferred to other generative design domains where spatial connectedness matters, such as fluid channel networks or vascular scaffolds.
  • The hierarchical feature-to-layer alignment principle may generalize beyond structural mechanics: any diffusion model conditioned on multi-scale physics data could benefit from matching feature abstraction to network depth.
  • Lightweight LoRA adaptation to non-square domains suggests that physics-guided pretraining produces transferable structural priors, reducing the data cost for new design domains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the hierarchical feature-depth correspondence is genuinely causal rather than architecture-specific, one would expect the optimal assignment to shift predictably when using architectures with different depth-to-scale relationships (e.g., transformer-based diffusion with explicit multi-scale tokens rather than UNet downsampling).
  • The FMS loss currently seeds heat from load positions only; for multi-load or multi-support problems, the choice of seed set becomes a design decision that could interact non-trivially with the connectivity penalty's behavior.
  • The orthogonality analysis showing U and PSL are highly correlated (score 0.20) yet contribute asymmetrically suggests the model may be exploiting representational differences rather than information differences; disentangling these would require controlled experiments with decorrelated features.
  • The claim that some generated designs outperform SIMP baselines in compliance could be tested more rigorously by comparing against diverse random restarts of SIMP rather than a single deterministic run, to distinguish genuine design-space exploration from noise in the FEA evaluation.

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 / 5 minor

Summary. This manuscript proposes HPG-Diff, a diffusion-based generative framework for topology optimization (TO). The method introduces two mechanisms: (1) a hierarchical physics-guided strategy that injects Displacement (U), Principal Stress Line (PSL), and Strain Energy Density (SED) features into shallow, mid, and deep UNet layers respectively via cross-attention; and (2) a Floating Material Suppression (FMS) loss, a differentiable connectivity constraint inspired by thermal conduction that penalizes disconnected material during training. The framework is evaluated on the TopoDiff dataset, demonstrating improvements in compliance error and floating material ratio over baselines including TopoDiff, DOM, TopologyGAN, NITO, and OAT. Case studies on non-square domains using LoRA fine-tuning are also presented.

Significance. The paper addresses a practical limitation of generative TO models—poor connectivity and generalizability—without relying on inference-time guidance or post-processing. The FMS loss is a novel, differentiable formulation of a load-seeded reachability penalty, which is a meaningful contribution over standard geometry-level connectivity losses. The hierarchical feature injection is conceptually motivated by the receptive field hierarchy of the UNet. The inclusion of sensitivity analyses (Appendix C), feature orthogonality discussion (Section 4.7.1), and comparisons with recent foundation models (OAT, NITO in Appendix G) adds rigor. The framework demonstrates strong quantitative performance on standard benchmarks.

major comments (3)
  1. The headline out-of-distribution (OOD) performance of 5.29% Avg CE (Abstract, Table 3) appears to be the best result among five inference trials rather than the expected value. Appendix D (Table D.14) reports five stochastic inference trials yielding a mean of 6.39%±1.33%. The value 5.29% is the minimum of these trials. This inflates the headline margin by approximately 21% relative to the seed-averaged mean. While the qualitative claim that HPG-Diff outperforms baselines holds even under the worst seed (8.32% vs. TopoDiff's 18.40%), the specific improvement margins cited in Section 4.4.3 (e.g., '71.25% reduction relative to TopoDiff-Guided') are computed against this optimistically selected number. The authors should clarify the sampling protocol for all methods (was it single-run, best-of-N, or seed-averaged?) and report the OOD results using the seed-averaged mean to ensure a fair and
  2. The long-tail failure cases significantly impact the OOD Avg CE. Table E.15 reports a maximum CE of 791.07% with a standard deviation of 31.76%, despite a median CE of 0.61%. This indicates that the mean is heavily skewed by a small number of catastrophic failures (2.70% of samples exceed 30% error). The high variance across seeds (1.33% in Table D.14) is likely driven by these long-tail cases. The manuscript should discuss these failure modes in the main text (not just the appendix) and clarify whether the FMS loss or hierarchical guidance can be strengthened to mitigate these severe disconnections, as they undermine the reliability of the expected performance.
  3. The justification for the specific layer-feature mapping (U to shallow, PSL to mid, SED to deep) relies heavily on a single visual example in Appendix B (Fig. B.9). While the orthogonality analysis in Section 4.7.1 shows U and PSL are highly correlated (score 0.2065), the authors argue they serve distinct roles. However, the quantitative evidence in Table B.10 is limited to one timestep sequence for one sample. To robustly support the claim that this specific hierarchical assignment is optimal and not overfit to the architecture, the authors should provide quantitative metrics (e.g., the M_X/M_bg ratio) averaged over a representative test set, or at least over multiple diverse examples.
minor comments (5)
  1. Section 4.4.3: The phrase 'reduces the key Average CE metric by 71.25% and 83.57% compared to TopoDiff-Guided and DOM w/ TA' should be re-evaluated in light of the seed-averaged OOD performance. The relative reduction against the mean (6.39%) would be approximately 65% and 80%, respectively.
  2. Table 8: The inference time for HPG-Diff is reported as 10.74s, which is slower than DOM (3.15s) and TopoDiff (6.95s). The text in Section 4.7.2 states the method is 'approximately twice as fast as the classical SIMP method' but does not explicitly address why it is slower than other generative baselines (TopoDiff, DOM) despite avoiding inference-phase guidance. A brief clarification on the preprocessing cost (5.92s for FEA) versus sampling cost (4.82s) would help contextualize this.
  3. Section 3.3, Eq. (6): The notation for the local maximum operator uses N_1(i,j), but the neighborhood definition is not explicitly stated (presumably a 3x3 grid). This should be clarified for reproducibility.
  4. Section 4.6: The case studies present designs that 'outperform the SIMP baseline in compliance.' It would be valuable to briefly note whether these generated designs were verified for manufacturability or if they contain thin members that might be problematic for fabrication, as this is a known issue in generative TO.
  5. Typo in Section 4.7.1: 'redundancy pattern among the physics features is stable under the evaluated unseen boundary conditions' — 'evaluated' could be omitted for conciseness.

Circularity Check

0 steps flagged

No significant circularity; the method is evaluated against external benchmarks and the physics features are derived from standard FEA on uniform density fields, not from the optimized target.

full rationale

The paper's central claims are evaluated against an external benchmark dataset (TopoDiff) and compared against independent baselines (TopologyGAN, TopoDiff, DOM, NITO, OAT). The physics features (U, PSL, SED) are computed from a single FEA solve on a uniform density field under prescribed boundary conditions and loads (Section 4.7.2), not from the optimized target topology, so the conditioning signals are not circularly derived from the generation target. The FMS loss (Section 3.3) is a parameter-free differentiable propagation process inspired by thermal conduction, with its hyperparameters (k, c, lambda_fm) validated via sensitivity analysis (Appendix C) rather than fitted to the test set. The hierarchical feature-to-layer assignment is justified by architectural reasoning (receptive field sizes) and tested via ablation (Appendix A), not assumed by definition. The orthogonality analysis (Section 4.7.1) is used to assess feature redundancy, not to define the features themselves. No self-citation chain is load-bearing for the central claims. The only minor concern is that the headline OOD number (5.29%) appears to be the best of five inference seeds (Appendix D, mean 6.39%±1.33%), but this is a reporting/selection issue, not a circularity in the derivation chain. The method's qualitative superiority over baselines holds even under the worst seed (8.32% vs. 18.40% for TopoDiff-Guided). Overall, the derivation is self-contained and independently falsifiable against external benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities or postulated objects. The free parameters (lambda_fm, c, k, tau) are standard hyperparameters selected via sensitivity analysis on the test set, which is a minor methodological concern but not circular. The domain assumptions are reasonable within the TO literature but the uniform-density physics feature assumption is the most fragile.

free parameters (4)
  • lambda_fm = 0.1
    Weight balancing denoising loss and FMS loss; selected via sensitivity analysis (Table C.13)
  • c = 5
    Scalar controlling exponential time-weighting w(t)=exp(-ct/T); selected via sensitivity analysis (Table C.12)
  • k = 64
    Propagation steps for FMS loss; set to domain dimension (64x64)
  • tau = 0.5
    Binarization threshold for evaluation; standard choice but affects reported metrics
axioms (4)
  • domain assumption UNet encoder-decoder hierarchy corresponds to local-to-global feature abstraction
    Section 3.2 assumes shallow layers handle local details while deep layers capture global context, justifying the layer-specific physics feature assignment
  • domain assumption Heat propagation from load positions is a valid proxy for structural connectivity in TO
    Section 3.3 assumes that material unreachable by heat propagation from the load is functionally disconnected and should be penalized
  • domain assumption Physics features computed on uniform density fields are informative for optimized topology
    Section 4.7.2 states FEA is solved on uniform initial density, assuming these fields contain sufficient information to guide generation of the optimized target
  • domain assumption Standard SIMP solutions are valid ground truth for training generative TO models
    Section 4.1 uses SIMP-generated dataset as ground truth, assuming SIMP produces optimal or near-optimal topologies

pith-pipeline@v1.1.0-glm · 25831 in / 2007 out tokens · 351093 ms · 2026-07-09T16:30:43.056256+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of HPG-Diff: Hierarchical physics-guided diffusion with differentiable connectivity constraints for topology optimization." pith.science (2026). https://pith.science/paper/PSXYRGKT

@misc{pith2026260707233,
  author       = {Pith},
  title        = {Pith review of: HPG-Diff: Hierarchical physics-guided diffusion with differentiable connectivity constraints for topology optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PSXYRGKT}},
  note         = {Machine review of arXiv:2607.07233}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Deep generative models offer a promising paradigm for topology optimization, enabling rapid design exploration. However, these approaches lack intrinsic physics guidance, often leading to poor generalizability across unseen boundary conditions and the formation of floating material artifacts. To address these limitations, we propose Hierarchical Physics-Guided Diffusion (HPG-Diff), a novel diffusion framework that enforces physics consistency through two synergistic mechanisms. First, we introduce a hierarchical physics-guided strategy that aligns different precomputed physics features with the denoising process, guiding material distribution toward optimal load paths to enhance generalizability. Second, we propose a floating material suppression loss as a differentiable connectivity constraint inspired by thermal conduction to improve topological connectivity. By simulating a virtual heat propagation process from load positions, this mechanism explicitly penalizes floating material during training. Quantitative evaluations demonstrate that HPG-Diff achieves average compliance errors of 0.87% (in-distribution) and 5.29% (out-of-distribution), while reducing floating material ratios to 2.90% and 2.44%, respectively. Furthermore, case studies on a 3:1 rectangular domain, including cantilever and bridge benchmarks, provide preliminary evidence that lightweight LoRA fine-tuning with a small dataset can support the adaptation of HPG-Diff to rectangular non-square domains.

Figures

Figures reproduced from arXiv: 2607.07233 by Boyuan Zhang, Jinbo Yang, Mingyue Yuan, Shikai Jing, Yoshifumi Kitamura.

Figure 1
Figure 1. Figure 1: Performance vs Efficiency: Engineering design demands multiple high-performing alternatives. Comparison of mainstream TO methods in adaptabil￾ity, performance, and computational cost. For a given TO design, strict inputs (boundary conditions, load) must always be satisfied, while relaxable inputs (volume fraction) allow adjustment, resulting in two modes of generating new designs. This trade-off is not onl… view at source ↗
Figure 2
Figure 2. Figure 2: Example of HPG-Diff for rapid and diverse topology optimization design generation. Given the boundary conditions and volume fraction, the model generates diverse and high-performing structures. Note that LoRA fine-tuning is employed to adapt the model to non-square design domains. 2. Related works 2.1. Physics-aware deep learning paradigms in topology optimization Learning-based computational methods have … view at source ↗
Figure 3
Figure 3. Figure 3: Example of local feature enhancements driven by conditioning physics features [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (HPG-Diff) Hierarchical Physics-Guided Diffusion Framework. Col￾ored lines depict how U (green), PSL (yellow), and SED (blue) hierarchically guide shal￾low, mid, and deep layers via cross-attention modules. these features enhance the UNet’s feature maps, improving the preservation of both details and overall structural integrity. Specifically, the corresponding conditioning physics features are: (1) Dis￾pl… view at source ↗
Figure 5
Figure 5. Figure 5: Negative (w/o FMS loss) vs Positive (w/ FMS loss) samples in [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Case study 1: Application of HPG-Diff in shelf bracket design. [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Case study 2: Application of HPG-Diff in cantilever beam design. [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Case study 3: Application of HPG-Diff in classic bridge design. [PITH_FULL_IMAGE:figures/full_fig_p022_8.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

69 extracted references · 69 canonical work pages · 4 internal anchors

  1. [1]

    Zhu, W.-H

    J.-H. Zhu, W.-H. Zhang, L. Xia, Topology optimization in aircraft and aerospace structures design, Archives of computational methods in en- gineering 23 (2016) 595–622

    work page 2016

  2. [2]

    Dagkolu, I

    A. Dagkolu, I. Gokdag, O. Yilmaz, Design and additive manufacturing of a fatigue-critical aerospace part using topology optimization and l-pbf process, Procedia Manufacturing 54 (2021) 238–243

    work page 2021

  3. [3]

    Guanghui, G

    S. Guanghui, G. Chengqi, Q. Dongliang, W. Dongtao, T. Lei, G. Tong, An aerospace bracket designed by thermo-elastic topology optimization and manufactured by additive manufacturing, Chinese Journal of Aero- nautics 33 (2020) 1252–1259

    work page 2020

  4. [4]

    M. N. Andersen, F. Wang, O. Sigmund, On the competition for ulti- 36 mately stiff and strong architected materials, Materials & Design 198 (2021) 109356

    work page 2021

  5. [5]

    Ferrari, O

    F. Ferrari, O. Sigmund, J. K. Guest, Topology optimization with lin- earized buckling criteria in 250 lines of matlab, Structural and Multi- disciplinary Optimization 63 (2021) 3045–3066

    work page 2021

  6. [6]

    J. M. Maestre, R. Hinchet, S. Coros, B. Thomaszewski, Toros: A topol- ogy optimization approach for designing robotic skins, ACM Transac- tions on Graphics (TOG) 42 (2023) 1–11

    work page 2023

  7. [7]

    M. Bi, P. Tran, L. Xia, G. Ma, Y. M. Xie, Topology optimization for 3d concrete printing with various manufacturing constraints, Additive Manufacturing 57 (2022) 102982

    work page 2022

  8. [8]

    W. Yang, L. Wang, G. Ma, P. Feng, An integrated method of topologi- cal optimization and path design for 3d concrete printing, Engineering Structures 291 (2023) 116435

    work page 2023

  9. [9]

    Ichihara, M

    N. Ichihara, M. Ueda, 3d-print infill generation using the biological phase field of an optimized discrete material orientation vector field, Composites Part B: Engineering 232 (2022) 109626

    work page 2022

  10. [10]

    García-Ávila, C

    J. García-Ávila, C. P. González-Gallegos, V. Segura-Ibarra, E. Vazquez, E. Garcia-Lopez, C. A. Rodríguez, A. Vargas-Martínez, E. Cuan- Urquizo, E. Ramírez-Cedillo, Dynamic topology optimization of 3d- printed transtibial orthopedic implant using tunable isotropic porous metamaterials, Journal of the Mechanical Behavior of Biomedical Ma- terials 153 (2024) 106479

    work page 2024

  11. [11]

    H. Li, B. Yao, Z. Li, Y. Peng, H. Fan, Compressive properties and de- formation mechanism of selective laser melting of ti6al4v porous femoral implants based on topological optimization, Composite Structures 321 (2023) 117326

    work page 2023

  12. [12]

    M.P.Bendsøe, Optimalshapedesignasamaterialdistributionproblem, Structural optimization 1 (1989) 193–202

    work page 1989

  13. [13]

    S. M. A. Seyed Mahmoud, G. Faraji, M. Baghani, M. S. Hashemi, A. Sheidaei, M. Baniassadi, Design of refractory alloys for desired ther- mal conductivity via ai-assisted in-silico microstructure realization, Ma- terials 16 (2023) 1088. 37

    work page 2023

  14. [14]

    J. Xu, J. Xu, Y. Sheng, T. Xie, Z. Wang, M. Elchalakani, Database- enabled machine learning framework for predicting fire performance in reinforced concrete columns: Focus on fire resistance, spalling behavior, and comparative evaluation against existing codes, Structures 88 (2026) 111794

    work page 2026

  15. [15]

    Y. Bu, Z. Cai, J. Peng, Y. Yu, J. Xu, Data-driven compressive strength estimation and mix optimization for ultra-high-performance concrete, Journal of Building Engineering (2025) 114253

    work page 2025

  16. [16]

    W. Chen, J. Xu, K. Yu, J. Yu, M. Ma, Z. Ma, Performance-based design of frp-confined recycled aggregate concrete powered by machine learning techniques, Engineering Structures 336 (2025) 120478

    work page 2025

  17. [17]

    Y. Wang, S. Zhang, Z. Zhang, Y. Yu, J. Xu, Reverse design for mixture proportions of recycled brick aggregate concrete using machine learning- based meta-heuristic algorithm: A multi-objective driven study, Journal of CO2 Utilization 88 (2024) 102944

    work page 2024

  18. [18]

    Zheng, S

    L. Zheng, S. Kumar, D. M. Kochmann, Data-driven topology opti- mization of spinodoid metamaterials with seamlessly tunable anisotropy, Computer Methods in Applied Mechanics and Engineering 383 (2021) 113894

    work page 2021

  19. [19]

    N. A. Kallioras, N. D. Lagaros, Dl-scale: a novel deep learning-based model order upscaling scheme for solving topology optimization prob- lems, Neural Computing and Applications 33 (2021) 7125–7144

    work page 2021

  20. [20]

    W. Qian, Y. Xu, H. Li, A topology description function-enhanced neural network for topology optimization, Computer-Aided Civil and Infras- tructure Engineering 38 (2023) 1020–1040

    work page 2023

  21. [21]

    J. Yan, D. Geng, Q. Xu, H. Li, Real-time topology optimization based on convolutional neural network by using retrain skill, Engineering with Computers 39 (2023) 4045–4059

    work page 2023

  22. [22]

    J. Yan, Q. Zhang, Q. Xu, Z. Fan, H. Li, W. Sun, G. Wang, Deep learning driven real time topology optimisation based on initial stress learning, Advanced Engineering Informatics 51 (2022) 101472. 38

    work page 2022

  23. [23]

    C. Deng, Y. Wang, C. Qin, Y. Fu, W. Lu, Self-directed online machine learning for topology optimization, Nature communications 13 (2022) 388

    work page 2022

  24. [24]

    J. Luo, Y. Li, W. Zhou, Z. Gong, Z. Zhang, W. Yao, An improved data- driven topology optimization method using feature pyramid networks with physical constraints., CMES-Computer Modeling in Engineering & Sciences 128 (2021)

    work page 2021

  25. [25]

    Jeong, J

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

    work page 2023

  26. [26]

    Jeong, C

    H. Jeong, C. Batuwatta-Gamage, J. Bai, C. Rathnayaka, Y. Zhou, Y. Gu, An advanced physics-informed neural network-based framework for nonlinear and complex topology optimization, Engineering Struc- tures 322 (2025) 119194

    work page 2025

  27. [27]

    Chandrasekhar, K

    A. Chandrasekhar, K. Suresh, Tounn: topology optimization using neu- ral networks, Structural and Multidisciplinary Optimization 63 (2021) 1135–1149

    work page 2021

  28. [28]

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio, Generative adversarial nets, Advances in neural information processing systems 27 (2014)

    work page 2014

  29. [29]

    J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, Advances in neural information processing systems 33 (2020) 6840–6851

    work page 2020

  30. [30]

    Regenwetter, A

    L. Regenwetter, A. H. Nobari, F. Ahmed, Deep generative models in engineering design: A review, Journal of Mechanical Design 144 (2022) 071704

    work page 2022

  31. [31]

    F. Mazé, F. Ahmed, Diffusion models beat gans on topology optimiza- tion, in: Proceedings of the AAAI conference on artificial intelligence, volume 37, 2023, pp. 9108–9116

    work page 2023

  32. [32]

    Giannone, A

    G. Giannone, A. Srivastava, O. Winther, F. Ahmed, Aligning optimiza- tion trajectories with diffusion models for constrained design generation, 39 Advances in Neural Information Processing Systems 36 (2023) 51830– 51861

    work page 2023

  33. [33]

    Zhang, G

    W. Zhang, G. Zhao, L. Su, Research on multi-stage topology optimiza- tion method based on latent diffusion model, Advanced Engineering Informatics 63 (2025) 102966

    work page 2025

  34. [34]

    Bayat, S

    Z. Bayat, S. Dindar, B. G. Erkal, Deep learning for post-earthquake road blockage and debris detection in resource-limited contexts, Journal of the Indian Society of Remote Sensing 54 (2026) 1257–1285

    work page 2026

  35. [35]

    Z. Li, P. Hilber, Z. Li, T. Laneryd, S. Ivanell, Temporally coordinated operation of green multi-energy airport microgrids with climatic correla- tions and flexible loads via decomposed stochastic programming, IEEE Transactions on Sustainable Energy 17 (2026) 1909–1922

    work page 2026

  36. [36]

    X. Zeng, X. Zheng, H. Chen, X. Yin, Z. Li, Y. Chen, Deriving the robust operating region for multistage dispatch of power systems: A decision- dependent uncertainties based modeling approach, IEEE Transactions on Sustainable Energy (2026) 1–12

    work page 2026

  37. [37]

    C. Feng, Z. Huang, J. Lin, L. Wang, Y. Zhang, F. Wen, Aggregation model and market mechanism for virtual power plant participation in inertia and primary frequency response, IEEE Transactions on Power Systems 41 (2026) 2101–2117

    work page 2026

  38. [38]

    C. Qiu, S. Du, J. Yang, A deep learning approach for efficient topology optimization based on the element removal strategy, Materials & Design 212 (2021) 110179

    work page 2021

  39. [39]

    C. Kim, J. Lee, J. Yoo, Machine learning-combined topology optimiza- tionforfunctionarygradedcompositestructuredesign, ComputerMeth- ods in Applied Mechanics and Engineering 387 (2021) 114158

    work page 2021

  40. [40]

    N. K. Brown, A. P. Garland, G. M. Fadel, G. Li, Deep reinforcement learning for engineering design through topology optimization of ele- mentally discretized design domains, Materials & Design 218 (2022) 110672. 40

    work page 2022

  41. [41]

    D. A. White, W. J. Arrighi, J. Kudo, S. E. Watts, Multiscale topology optimizationusingneuralnetworksurrogatemodels, ComputerMethods in Applied Mechanics and Engineering 346 (2019) 1118–1135

    work page 2019

  42. [42]

    Z. Nie, T. Lin, H. Jiang, L. B. Kara, Topologygan: Topology optimiza- tion using generative adversarial networks based on physical fields over the initial domain, Journal of Mechanical Design 143 (2021) 031715

    work page 2021

  43. [43]

    Zhang, A

    L. Zhang, A. Rao, M. Agrawala, Adding conditional control to text-to- image diffusion models, in: Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 3836–3847

    work page 2023

  44. [44]

    C. Mou, X. Wang, L. Xie, Y. Wu, J. Zhang, Z. Qi, Y. Shan, T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models, in: Proceedings of the AAAI conference on artificial intelligence, volume 38, 2024, pp. 4296–4304

    work page 2024

  45. [45]

    C. Zeng, Y. Dong, P. Peers, Y. Kong, H. Wu, X. Tong, Dilightnet: Fine-grained lighting control for diffusion-based image generation, in: ACM SIGGRAPH 2024 Conference Papers, 2024, pp. 1–12

    work page 2024

  46. [46]

    Y. Li, G. Baciu, Sg-gan: Adversarial self-attention gcn for point cloud topological parts generation, IEEE Transactions on Visualization and Computer Graphics 28 (2021) 3499–3512

    work page 2021

  47. [47]

    Y. Li, H. Liu, Q. Wu, F. Mu, J. Yang, J. Gao, C. Li, Y. J. Lee, Gli- gen: Open-set grounded text-to-image generation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 22511–22521

    work page 2023

  48. [48]

    Dhariwal, A

    P. Dhariwal, A. Nichol, Diffusion models beat gans on image synthesis, Advances in neural information processing systems 34 (2021) 8780–8794

    work page 2021

  49. [49]

    J. Ho, T. Salimans, Classifier-free diffusion guidance, in: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021

    work page 2021

  50. [50]

    Voynov, K

    A. Voynov, K. Aberman, D. Cohen-Or, Sketch-guided text-to-image diffusion models, in: ACM SIGGRApH 2023 conference proceedings, 2023, pp. 1–11. 41

    work page 2023

  51. [51]

    Sigmund, K

    O. Sigmund, K. Maute, Topology optimization approaches: A compar- ative review, Structural and multidisciplinary optimization 48 (2013) 1031–1055

    work page 2013

  52. [52]

    J. D. Deaton, R. V. Grandhi, A survey of structural and multidisci- plinary continuum topology optimization: post 2000, Structural and Multidisciplinary Optimization 49 (2014) 1–38

    work page 2000

  53. [53]

    V. Cool, N. Aage, O. Sigmund, A practical review on promoting con- nectivity in topology optimization, Structural and Multidisciplinary Optimization 68 (2025) 73

    work page 2025

  54. [54]

    Patel, D

    D. Patel, D. Bielecki, R. Rai, G. Dargush, Improving connectivity and accelerating multiscale topology optimization using deep neural network techniques, Structural and Multidisciplinary Optimization 65 (2022) 126

    work page 2022

  55. [55]

    M. M. Behzadi, H. T. Ilieş, Gantl: Toward practical and real-time topology optimization with conditional generative adversarial networks and transfer learning, Journal of Mechanical Design 144 (2022) 021711

    work page 2022

  56. [56]

    E. J. Hu, yelong shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, LoRA: Low-rank adaptation of large language models, in: International Conference on Learning Representations, 2022

    work page 2022

  57. [57]

    Mosinska, P

    A. Mosinska, P. Marquez-Neila, M. Koziński, P. Fua, Beyond the pixel- wise loss for topology-aware delineation, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 3136– 3145

    work page 2018

  58. [58]

    X. Hu, F. Li, D. Samaras, C. Chen, Topology-preserving deep image segmentation, Advances in neural information processing systems 32 (2019)

    work page 2019

  59. [59]

    S. Shit, J. C. Paetzold, A. Sekuboyina, I. Ezhov, A. Unger, A. Zhylka, J. P. Pluim, U. Bauer, B. H. Menze, cldice-a novel topology-preserving loss function for tubular structure segmentation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 16560–16569. 42

    work page 2021

  60. [60]

    Crane, C

    K. Crane, C. Weischedel, M. Wardetzky, Geodesics in heat: A new approach to computing distance based on heat flow, ACM Transactions on Graphics (ToG) 32 (2013) 1–11

    work page 2013

  61. [61]

    Decoupled Weight Decay Regularization

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, arXiv preprint arXiv:1711.05101 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  62. [62]

    D. Luo, J. Wu, J. Wang, H. Xie, X. Yue, S. Tang, Difffluid: Plain diffusion models are effective predictors of flow dynamics, arXiv preprint arXiv:2409.13665 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  63. [63]

    J. Qiu, J. Huang, X. Zhang, Z. Lin, M. Pan, Z. Liu, F. Miao, Pi-fusion: Physics-informed diffusion model for learning fluid dynamics, arXiv preprint arXiv:2406.03711 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  64. [64]

    S. Deng, Q. Zhang, Z. Cai, Generative method for aerodynamic opti- mization based on classifier-free guided denoising diffusion probabilistic model, arXiv preprint arXiv:2503.07056 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  65. [65]

    Bhattacharya, L

    K. Bhattacharya, L. Cao, A. Stuart, Optimal experimental design for re- liable learning of history-dependent constitutive laws, Computer Meth- ods in Applied Mechanics and Engineering 457 (2026) 119022

    work page 2026

  66. [66]

    M. A. Seyed Mahmoud, D. Renner, A. Khosravani, S. R. Kalidindi, Sequential bayesian inference of the gtn damage model using multimodal experimental data, Acta Materialia 306 (2026) 121902

    work page 2026

  67. [67]

    T.-H. Kwok, Y. Li, Y. Chen, A structural topology design method based on principal stress line, Computer-Aided Design 80 (2016) 19–31

    work page 2016

  68. [68]

    A. H. Nobari, L. Regenwetter, G. Giannone, F. Ahmed, NITO: Neural implicit fields for resolution-free and domain-adaptable topology opti- mization, Transactions on Machine Learning Research (2025)

    work page 2025

  69. [69]

    Heyrani Nobari, L

    A. Heyrani Nobari, L. Regenwetter, C. Picard, L. Han, F. Ahmed, Op- timize any topology: A foundation model for shape-and resolution-free structural topology optimization, Advances in Neural Information Pro- cessing Systems 38 (2025) 153374–153419. 43

    work page 2025