pith. sign in

arxiv: 2605.19110 · v1 · pith:46M3VERVnew · submitted 2026-05-18 · 💻 cs.CE

IterSIMP-{σ}: Evaluating LLM-Assisted Spatial Interventions in Stress-Aware Topology Optimization

Shaoliang Yang , Jun Wang , Yunsheng Wang This is my paper

Pith reviewed 2026-05-20 07:25 UTC · model grok-4.3

classification 💻 cs.CE
keywords topology optimizationLLM-assisted designstress-aware optimizationspatial interventionsSIMP methodmultimodal LLMsdesign automationdensity seeding
0
0 comments X

The pith

The IterSIMP-σ system supports multimodal LLMs as inspectable spatial proposal modules inside stress-aware topology optimization.

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

This paper evaluates whether multimodal LLMs can serve as spatial proposal modules for stress-aware topology optimization while leaving the core SIMP optimizer unchanged. The IterSIMP-σ loop renders density and von Mises stress fields after each solve, feeds them to a hybrid LLM/rule interpreter that ranks spatial interventions, and lets deterministic gates accept, reject, or stop the proposals. The main intervention is a soft density seed that raises the initial density of selected elements for the next optimality-criteria update but leaves them free to evolve. On a 16-problem 2D benchmark the LLM condition produces a small compliance difference that is not statistically significant, while a fixed-volume study shows 25 of 44 completed LLM runs yielding all-gate-passing feasible states. A sympathetic reader would care because the design keeps proposals reviewable and separates the AI module from the physics solver, offering a modular path for AI assistance in engineering optimization.

Core claim

The paper claims that IterSIMP-σ functions as an inspectable LLM-assisted design-automation framework for spatial interventions. The central mechanism renders density and stress fields after each SIMP solve, employs a hybrid LLM/rule interpreter to propose ranked spatial interventions as soft density seeds, and applies deterministic gates to control which proposals advance. In the fixed-volume attribution study the LLM condition completed 44 of 48 evaluations, with 25 of 44 producing all-gate-passing retained states and a mean normalized seed-to-hotspot distance of 0.221 for accepted actions. The 2D controller-policy benchmark yields a 1.9 percent lower geometric mean compliance for the soft

What carries the argument

The IterSIMP-σ loop that renders density and von Mises stress fields, uses a hybrid LLM/rule interpreter to propose ranked spatial interventions, and applies deterministic gates to accept, reject, or stop soft density seed actions.

If this is right

  • Soft density seeding lets LLM proposals influence the optimality-criteria update without locking in hard constraints.
  • Hybrid LLM and rule-based control produces ranked proposals that remain reviewable at each optimization step.
  • The separation of the compliance solver from the spatial proposal module allows direct comparison of LLM proposals against deterministic max-stress hotspot seeding and random stress-region seeding.
  • Accepted LLM actions show a mean normalized distance of 0.221 from stress hotspots, indicating targeted rather than random interventions.
  • The framework extends naturally to 3D problems with the same structure of render-interpret-gate steps.

Where Pith is reading between the lines

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

  • If scaled to more complex loads or manufacturing constraints, the inspectable proposals could let engineers inject domain knowledge through visual or language prompts.
  • Similar hybrid loops might apply to other physics-driven optimizations such as thermal or fluid problems where stress-like fields can be visualized.
  • A larger benchmark with statistical power could test whether the current small compliance gap becomes significant when volume is strictly fixed across all methods.

Load-bearing premise

The 16-problem 2D controller-policy benchmark and six-problem 3D extension are representative enough to evaluate whether LLM spatial proposals meaningfully advance stress-aware optimization.

What would settle it

A larger benchmark set that compares fixed-volume feasible-final compliance between LLM-proposed soft seeds and deterministic exact-hotspot seeding and finds no consistent advantage for the LLM condition.

Figures

Figures reproduced from arXiv: 2605.19110 by Jun Wang, Shaoliang Yang, Yunsheng Wang.

Figure 1
Figure 1. Figure 1: Schematic overview of the benchmark IterSIMP- [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Passive-solid seeding (left) vs. soft density seeding (right) for the cantilever-with-two-voids [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Retained-best compliance ratio for all 16 two-dimensional controller-policy comparisons at [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative seed-42 density fields for the five two-dimensional benchmark problems classified [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Representative seed-42 density fields for the two-dimensional benchmark cases that are practical [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Representative seed-42 density fields for the three cases where the rule-based run has a lower [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Fixed-volume 2D primary endpoints. Panel A decomposes attempted condi [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Spatial-action localization analysis across fixed-volume policies. Panel A shows per-action [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Stress-satisfaction rate vs. σyield calibration percentile. The deterministic calibration study uses the rule-based interpreter path and therefore coincides with the matched rule study. This figure is calibration evidence, not controller-comparison evidence [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Per-problem retained-compliance sensitivity to the [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Step-by-step compliance convergence for seed-42 traces ( [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Controller sequence for the L-bracket in the primary seed-42 evaluation run ( [PITH_FULL_IMAGE:figures/full_fig_p028_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Representative seed-42 density and von Mises stress fields for the Michell truss ( [PITH_FULL_IMAGE:figures/full_fig_p029_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: MBB beam seed sensitivity (CV = 0.01%). All three soft-seed repeats are nearly identical (C = 154.48, 154.47, 154.52), producing slightly lower rule-based retained compliance rather than a soft-seed advantage [PITH_FULL_IMAGE:figures/full_fig_p030_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Low-volume cantilever seed sensitivity (CV = 4.5%). The three soft-seed repeats pass all seven gate checks with C = 105.83, 95.18, and 98.02 [PITH_FULL_IMAGE:figures/full_fig_p030_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Michell truss seed sensitivity (CV = 12.1%). The three soft-seed repeats give C = 29.80, 26.25, and 22.08; the first final state fails gate checks and the latter two pass all seven gate checks. Figures 14–16 together constitute the seed sensitivity analysis (LLM condition; density top row, von Mises stress bottom row, seeds 42/123/7 left to right). 30 [PITH_FULL_IMAGE:figures/full_fig_p030_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Isosurface renderings (ρe = 0.5 threshold) and von Mises stress fields for the four non-tie 3D benchmark problems. Images are seed-42 examples; compliance percentages use three-seed mean retained-best compliance Crep. Left two columns: LLM (blue border); right two columns: rule-based (gray border). Stress colors are normalized per panel and are qualitative within-panel visualizations only; the 3D bridge L… view at source ↗
Figure 18
Figure 18. Figure 18: Representative seed-42 density fields and density slices through mid-planes (X, Y, Z) for the [PITH_FULL_IMAGE:figures/full_fig_p033_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Three-way ablation summarized as retained-compliance ratios for all 16 two-dimensional [PITH_FULL_IMAGE:figures/full_fig_p034_19.png] view at source ↗
read the original abstract

This paper studies whether multimodal large language models (LLMs) can serve as inspectable spatial proposal modules for stress-aware topology optimization. IterSIMP-{\sigma} keeps the SIMP optimizer as a compliance-minimizing finite-element solver and places a deterministic stress pass, gate evaluator, and hybrid LLM/rule interpreter around it. After each solve, density and von Mises stress fields are rendered; the interpreter proposes ranked spatial interventions; and deterministic safeguards accept, reject, or stop each action. The main action is a soft density seed, where selected elements are initialized at elevated density before the next solve but remain free under the optimality-criteria update. We evaluate the loop on a 16-problem 2D controller-policy benchmark, a six-problem exploratory 3D extension, passive-solid and input ablations, stress-threshold sensitivity, and a fixed-volume attribution study comparing LLM proposals with deterministic max-stress hotspot seeding, random stress-region seeding, and rule-based control. The 2D controller-policy benchmark shows a small retained-compliance difference (1.9% lower geometric mean for the soft-seed LLM), but this diagnostic is not statistically significant (W = 33, two-sided p = 0.382) and is not a fixed-volume feasible-final comparison. In the fixed-volume study, the LLM condition completed 44/48 attempted evaluations; 25/44 completed evaluations produced all-gate-passing retained states. Feasible-final scoring against rule-based control is split 4/4/1, and deterministic exact-hotspot seeding remains competitive. Accepted LLM spatial actions with per-step records have mean normalized seed-to-hotspot distance 0.221. The results support IterSIMP-{\sigma} as an inspectable LLM-assisted design-automation framework for spatial interventions, not yet as evidence that LLM visual reasoning improves stress-constrained optimization.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces IterSIMP-σ, a hybrid framework that retains the deterministic SIMP optimizer for compliance minimization while wrapping it with multimodal LLM proposals for spatial interventions (e.g., soft density seeding) after rendering density and von Mises stress fields. Deterministic stress passes and gate evaluators accept, reject, or halt actions. Evaluation covers a 16-problem 2D controller-policy benchmark (1.9% geometric-mean compliance advantage, Wilcoxon W=33, p=0.382, non-significant), a 6-problem 3D extension, ablations, and a fixed-volume attribution study (44/48 completions, 25/44 all-gate-passing states, 4/4/1 split vs. rule-based control, mean normalized seed-to-hotspot distance 0.221). The abstract concludes that results support the framework as inspectable but not yet as evidence of LLM improvement over stress-constrained optimization.

Significance. If the framework claim holds, the work provides a concrete, safeguarded example of LLM integration into established topology optimization, emphasizing inspectability through per-step records and deterministic gates. Strengths include explicit baselines (rule-based control, random stress-region seeding, deterministic max-stress hotspot seeding) and statistical testing (Wilcoxon), which allow direct comparison rather than self-referential evaluation. This could inform hybrid AI-traditional design automation in computational engineering, though the modest quantitative gains position the contribution as primarily methodological and proof-of-concept.

major comments (2)
  1. Abstract and evaluation sections: The support for IterSIMP-σ as an inspectable LLM-assisted framework rests on the fixed-volume attribution study (25/44 all-gate-passing states, 4/4/1 feasible-final split vs. rule-based control), yet deterministic exact-hotspot seeding remains competitive and LLM proposals show mean normalized seed-to-hotspot distance of 0.221; this indicates limited added value beyond simple heuristics and weakens the framework-utility claim.
  2. 2D controller-policy benchmark (evaluation sections): The 16-problem benchmark reports a non-significant compliance difference (W=33, p=0.382) that is explicitly not a fixed-volume feasible-final comparison, and the six-problem 3D extension is exploratory; with these scales and outcomes, the evidence does not yet securely establish even the modest claim of framework support for spatial interventions in stress-aware optimization.
minor comments (1)
  1. Abstract: The 44/48 completion rate would benefit from explicit clarification on whether it aggregates across all conditions or applies only to the LLM condition.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We agree that the reported performance differences are modest and non-significant, and the manuscript already qualifies its claims accordingly, positioning the work as methodological support for an inspectable hybrid framework rather than evidence of LLM superiority. We respond to each major comment below.

read point-by-point responses

  1. Referee: Abstract and evaluation sections: The support for IterSIMP-σ as an inspectable LLM-assisted framework rests on the fixed-volume attribution study (25/44 all-gate-passing states, 4/4/1 feasible-final split vs. rule-based control), yet deterministic exact-hotspot seeding remains competitive and LLM proposals show mean normalized seed-to-hotspot distance of 0.221; this indicates limited added value beyond simple heuristics and weakens the framework-utility claim.

    Authors: We thank the referee for this observation. The fixed-volume attribution study is intended to demonstrate inspectability, gate reliability, and per-step transparency of LLM proposals rather than performance gains. The mean normalized seed-to-hotspot distance of 0.221 indicates proposals informed by the full rendered stress and density fields, not merely the single maximum point. While exact-hotspot seeding is competitive, the framework's contribution is the safeguarded, recordable integration that permits analysis of accepted interventions. The 25/44 all-gate-passing rate and 4/4/1 feasible-final split provide evidence that LLM spatial actions can be incorporated reliably. We do not revise the claims, as the abstract already states the results support the framework as inspectable but not yet as improvement over stress-constrained optimization. revision: no

  2. Referee: 2D controller-policy benchmark (evaluation sections): The 16-problem benchmark reports a non-significant compliance difference (W=33, p=0.382) that is explicitly not a fixed-volume feasible-final comparison, and the six-problem 3D extension is exploratory; with these scales and outcomes, the evidence does not yet securely establish even the modest claim of framework support for spatial interventions in stress-aware optimization.

    Authors: We agree the 2D compliance difference is non-significant and is presented in the manuscript as a diagnostic metric, not a fixed-volume feasible-final comparison. The modest claim of framework support rests primarily on the attribution study, which supplies detailed per-step records and deterministic gate outcomes showing that spatial interventions can be safely integrated. The 3D extension is explicitly exploratory. With the given scales we recognize the limits on statistical strength, yet the current evidence substantiates feasibility and inspectability of the hybrid loop. We will add a clarifying sentence in the discussion reiterating the diagnostic role of the benchmarks and the exploratory status of the 3D results. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical evaluation rests on external baselines and statistical tests

full rationale

The paper describes an experimental framework IterSIMP-σ that wraps a standard SIMP optimizer with LLM spatial proposals, deterministic stress gates, and rule-based interpreters. All reported outcomes derive from direct comparisons on fixed 16-problem 2D and 6-problem 3D benchmark sets against independent deterministic controls (exact-hotspot seeding, random stress-region seeding, rule-based control) plus Wilcoxon signed-rank tests and gate-pass counts. No equations, fitted parameters, or predictions are defined in terms of the target quantities; the 1.9 % compliance difference and 25/44 all-gate-passing states are measured quantities, not self-generated by construction. Self-citations are absent from the load-bearing claims, and the evaluation is therefore self-contained against external methods.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard topology-optimization assumptions plus the new hybrid interpreter component; no new physical entities are postulated and free parameters are limited to implementation choices such as stress thresholds and prompt design.

free parameters (2)
  • stress-threshold value
    Used inside the gate evaluator to decide acceptance of LLM proposals; value chosen to balance safety and intervention rate.
  • LLM prompt wording and temperature
    Choices that control how density and stress images are described to the model and how ranked interventions are requested.
axioms (2)
  • domain assumption SIMP optimality-criteria update correctly minimizes compliance subject to volume constraint
    Invoked when the soft-seeded elements remain free under the standard update rule.
  • domain assumption Rendered density and von Mises stress fields are faithful visual inputs for LLM reasoning
    Required for the multimodal interpreter to produce spatially relevant proposals.

pith-pipeline@v0.9.0 · 5878 in / 1579 out tokens · 37406 ms · 2026-05-20T07:25:36.862402+00:00 · methodology

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

49 extracted references · 49 canonical work pages · 3 internal anchors

  1. [1]

    Bendsøe and Ole Sigmund.Topology Optimization

    Martin P. Bendsøe and Ole Sigmund.Topology Optimization. Springer, Berlin, 2004. doi: 10.1007/978-3-662-05086-6

    work page doi:10.1007/978-3-662-05086-6 2004

  2. [2]

    Topology optimization approaches.Structural and Multi- disciplinary Optimization, 48(6):1031–1055, 2013

    Ole Sigmund and Kurt Maute. Topology optimization approaches.Structural and Multi- disciplinary Optimization, 48(6):1031–1055, 2013. doi: 10.1007/s00158-013-0978-6

    work page doi:10.1007/s00158-013-0978-6 2013

  3. [3]

    Bendsøe and Noboru Kikuchi

    Martin P. Bendsøe and Noboru Kikuchi. Generating optimal topologies in structural design usingahomogenizationmethod.Computer Methods in Applied Mechanics and Engineering, 71(2):197–224, 1988. doi: 10.1016/0045-7825(88)90086-2

    work page doi:10.1016/0045-7825(88)90086-2 1988

  4. [4]

    A 99 line topology optimization code written in Matlab.Structural and Multidisciplinary Optimization, 21(2):120–127, 2001

    Ole Sigmund. A 99 line topology optimization code written in Matlab.Structural and Multidisciplinary Optimization, 21(2):120–127, 2001. doi: 10.1007/s001580050176

    work page doi:10.1007/s001580050176 2001

  5. [5]

    Lazarov, and Ole Sig- mund

    Erik Andreassen, Anders Clausen, Mattias Schevenels, Boyan S. Lazarov, and Ole Sig- mund. Efficient topology optimization in MATLAB using 88 lines of code.Structural and Multidisciplinary Optimization, 43(1):1–16, 2011. doi: 10.1007/s00158-010-0594-7

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

  6. [6]

    Lazarov, and Ole Sigmund

    Fengwen Wang, Boyan S. Lazarov, and Ole Sigmund. On projection methods, conver- gence and robust formulations in topology optimization.Structural and Multidisciplinary Optimization, 43(6):767–784, 2011. doi: 10.1007/s00158-010-0602-y

    work page doi:10.1007/s00158-010-0602-y 2011

  7. [7]

    Lazarov and Ole Sigmund

    Boyan S. Lazarov and Ole Sigmund. Filters in topology optimization based on Helmholtz- type differential equations.International Journal for Numerical Methods in Engineering, 86(6):765–781, 2011. doi: 10.1002/nme.3072

    work page doi:10.1002/nme.3072 2011

  8. [8]

    Numerical instabilities in topology optimization: a survey on procedures dealing with checkerboards, mesh-dependencies and local minima

    Ole Sigmund and Joakim Petersson. Numerical instabilities in topology optimization: a survey on procedures dealing with checkerboards, mesh-dependencies and local minima. Structural Optimization, 16(1):68–75, 1998. doi: 10.1007/BF01214002

    work page doi:10.1007/bf01214002 1998

  9. [9]

    Pierre Duysinx and Martin P. Bendsøe. Topology optimization of continuum struc- tures with local stress constraints.International Journal for Numerical Methods in En- gineering, 43(8):1453–1478, 1998. doi: 10.1002/(SICI)1097-0207(19981230)43:8<1453:: AID-NME480>3.0.CO;2-2

    work page doi:10.1002/(sici)1097-0207(19981230)43:8 1998

  10. [10]

    Stress-based topology optimization for continua.Structural and Multidisciplinary Optimization, 41(4): 605–620, 2010

    Chau Le, Julian Norato, Tyler Bruns, Christopher Ha, and Daniel Tortorelli. Stress-based topology optimization for continua.Structural and Multidisciplinary Optimization, 41(4): 605–620, 2010. doi: 10.1007/s00158-009-0440-y. 41

    work page doi:10.1007/s00158-009-0440-y 2010

  11. [11]

    R. J. Yang and C. J. Chen. Stress-based topology optimization.Structural Optimization, 12(2–3):98–105, 1996. doi: 10.1007/BF01196941

    work page doi:10.1007/bf01196941 1996

  12. [12]

    Stress constrained topology optimization.Structural and Multidisciplinary Optimization, 48(1):33–47, 2013

    Erik Holmberg, Börje Torstenfelt, and Anders Klarbring. Stress constrained topology optimization.Structural and Multidisciplinary Optimization, 48(1):33–47, 2013. doi: 10. 1007/s00158-012-0880-7

    work page 2013

  13. [13]

    Senhora, Oliver Giraldo-Londoño, Ivan F

    Fernando V. Senhora, Oliver Giraldo-Londoño, Ivan F. M. Menezes, and Glaucio H. Paulino. Topology optimization with local stress constraints: a stress aggregation-free approach.Structural and Multidisciplinary Optimization, 62(4):1639–1668, 2020. doi: 10.1007/s00158-020-02573-9

    work page doi:10.1007/s00158-020-02573-9 2020

  14. [14]

    Beck, and Ole Sigmund

    Gustavo Assis da Silva, Niels Aage, André T. Beck, and Ole Sigmund. Three-dimensional manufacturing tolerant topology optimization with hundreds of millions of local stress constraints.International Journal for Numerical Methods in Engineering, 122(2):548–578,

    work page

  15. [15]

    doi: 10.1002/nme.6548

    work page doi:10.1002/nme.6548

  16. [16]

    Senhora, Ivan F

    Fernando V. Senhora, Ivan F. M. Menezes, and Glaucio H. Paulino. Topology optimiza- tion with local stress constraints and continuously varying load direction and magnitude: towards practical applications.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 479(2271):20220436, 2023. doi: 10.1098/rspa.2022.0436

    work page doi:10.1098/rspa.2022.0436 2023

  17. [17]

    Stress-constrained topology optimiza- tion with the augmented lagrangian method: A comparative study of subproblem solvers

    Gustavo Assis Da Silva and Hélio Emmendoerfer Jr. Stress-constrained topology optimiza- tion with the augmented lagrangian method: A comparative study of subproblem solvers. International Journal for Numerical Methods in Engineering, 126(12):e70066, 2025. doi: 10.1002/nme.70066

    work page doi:10.1002/nme.70066 2025

  18. [18]

    Topology optimization: a review for structural designs under statics problems

    Tianshu Tang, Leijia Wang, Mingqiao Zhu, Huzhi Zhang, Jiarui Dong, Wenhui Yue, and Hui Xia. Topology optimization: a review for structural designs under statics problems. Materials, 17(23):5970, 2024. doi: 10.3390/ma17235970

    work page doi:10.3390/ma17235970 2024

  19. [19]

    Topology optimization for minimum weight with com- pliance and stress constraints.Structural and Multidisciplinary Optimization, 46(3):369– 384, 2012

    Matteo Bruggi and Pierre Duysinx. Topology optimization for minimum weight with com- pliance and stress constraints.Structural and Multidisciplinary Optimization, 46(3):369– 384, 2012. doi: 10.1007/s00158-012-0759-7

    work page doi:10.1007/s00158-012-0759-7 2012

  20. [20]

    Oliver Giraldo-Londoño and Glaucio H. Paulino. A unified approach for topology optimiza- tion with local stress constraints considering various failure criteria: von Mises, Drucker– Prager, Tresca, Mohr–Coulomb, Bresler–Pister and Willam–Warnke.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 476(2238):20190861,

    work page

  21. [21]

    doi: 10.1098/rspa.2019.0861

    work page doi:10.1098/rspa.2019.0861 2019

  22. [22]

    Deep learning for determining a near-optimal topological design without any iteration.Structural and Multidisciplinary Optimization, 59(3):787–799, 2019

    Yonggyun Yu, Taeil Hur, Jaeho Jung, and In Gwun Jang. Deep learning for determining a near-optimal topological design without any iteration.Structural and Multidisciplinary Optimization, 59(3):787–799, 2019. doi: 10.1007/s00158-018-2101-5

    work page doi:10.1007/s00158-018-2101-5 2019

  23. [23]

    Neural networks for topology optimization.Russian Journal of Numerical Analysis and Mathematical Modelling, 34(4):215–223, 2019

    Ivan Sosnovik and Ivan Oseledets. Neural networks for topology optimization.Russian Journal of Numerical Analysis and Mathematical Modelling, 34(4):215–223, 2019. doi: 10.1515/rnam-2019-0018

    work page doi:10.1515/rnam-2019-0018 2019

  24. [24]

    TopologyGAN: Topology optimization using generative adversarial networks based on physical fields over the initial domain.Journal of Mechanical Design, 143(3):031715, 2021

    Zhenguo Nie, Tong Lin, Haoliang Jiang, and Levent Burak Kara. TopologyGAN: Topology optimization using generative adversarial networks based on physical fields over the initial domain.Journal of Mechanical Design, 143(3):031715, 2021. doi: 10.1115/1.4049533

    work page doi:10.1115/1.4049533 2021

  25. [25]

    Self-directed online machine learning for topology optimization.Nature Communications, 13(1):388, 2022

    Changyu Deng, Yizhou Wang, Can Qin, Yun Fu, and Wei Lu. Self-directed online machine learning for topology optimization.Nature Communications, 13(1):388, 2022. doi: 10. 1038/s41467-021-27713-7. 42

    work page 2022

  26. [26]

    Algorithmically-consistent deep learning frameworks for struc- tural topology optimization.Engineering Applications of Artificial Intelligence, 106:104483,

    Jaydeep Rade, Aditya Balu, Ethan Herron, Jay Pathak, Rishikesh Ranade, Soumik Sarkar, and Adarsh Krishnamurthy. Algorithmically-consistent deep learning frameworks for struc- tural topology optimization.Engineering Applications of Artificial Intelligence, 106:104483,

    work page

  27. [27]

    doi: 10.1016/j.engappai.2021.104483

    work page doi:10.1016/j.engappai.2021.104483 2021

  28. [28]

    Topology optimization via machine learning and deep learning: a review.Journal of Computational Design and Engineering, 10(4):1736–1766, 2023

    Seungyeon Shin, Dongju Shin, and Namwoo Kang. Topology optimization via machine learning and deep learning: a review.Journal of Computational Design and Engineering, 10(4):1736–1766, 2023. doi: 10.1093/jcde/qwad072

    work page doi:10.1093/jcde/qwad072 2023

  29. [29]

    Yiming Zhang, Chen Jia, Hongyi Zhang, Naiyu Fang, Shuyou Zhang, and Nam H. Kim. Improving data-efficiency of deep generative model for fast design synthesis.Journal of Me- chanical Science and Technology, 38(4):1957–1970, 2024. doi: 10.1007/s12206-024-0328-1

    work page doi:10.1007/s12206-024-0328-1 1957

  30. [30]

    Topology optimization using neural networks with conditioning field initialization for improved efficiency.Journal of Mechanical Design, 146(6):061702, 2024

    Hongrui Chen, Aditya Joglekar, and Levent Burak Kara. Topology optimization using neural networks with conditioning field initialization for improved efficiency.Journal of Mechanical Design, 146(6):061702, 2024. doi: 10.1115/1.4064131

    work page doi:10.1115/1.4064131 2024

  31. [31]

    Kallioras, Georgios Kazakis, and Nikos D

    Nikos A. Kallioras, Georgios Kazakis, and Nikos D. Lagaros. Accelerated topology opti- mization by means of deep learning.Structural and Multidisciplinary Optimization, 62(3): 1185–1212, 2020. doi: 10.1007/s00158-020-02545-z

    work page doi:10.1007/s00158-020-02545-z 2020

  32. [32]

    GPT-4 Technical Report

    OpenAI. GPT-4 technical report, 2023. arXiv preprint arXiv:2303.08774

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  33. [33]

    Gemini: A Family of Highly Capable Multimodal Models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, et al. Gemini: A family of highly capable multimodal models, 2023. arXiv preprint arXiv:2312.11805

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  34. [34]

    How can large language models help humans in design and manufacturing? Part 1: Elements of the LLM-enabled computational design and manufacturing pipeline

    Liane Makatura, Michael Foshey, Bohan Wang, Felix Hähnlein, Pingchuan Ma, Bolei Deng, Megan Tjandrasuwita, Andrew Spielberg, Crystal Owens, Peter Yichen Chen, Allan Zhao, Amy Zhu, Wil Norton, Edward Gu, Joshua Jacob, Yifei Li, Adriana Schulz, and Wojciech Matusik. How can large language models help humans in design and manufacturing? Part 1: Elements of t...

    work page 2024

  35. [35]

    Sizhong Qin, Hong Guan, Wenjie Liao, Yi Gu, Zhe Zheng, Hongjing Xue, and Xinzheng Lu. Intelligent design and optimization system for shear wall structures based on large language models and generative artificial intelligence.Journal of Building Engineering, 95: 109996, 2024. doi: 10.1016/j.jobe.2024.109996

    work page doi:10.1016/j.jobe.2024.109996 2024

  36. [36]

    Integrating large language models for automated structural analysis, 2025

    Haoran Liang, Mohammad Talebi Kalaleh, and Qipei Mei. Integrating large language models for automated structural analysis, 2025. arXiv preprint arXiv:2504.09754

    work page arXiv 2025

  37. [37]

    OR-LLM- Agent: Automating modeling and solving of operations research optimization problems with reasoning LLM, 2025

    Bowen Zhang, Pengcheng Luo, Genke Yang, Boon-Hee Soong, and Chau Yuen. OR-LLM- Agent: Automating modeling and solving of operations research optimization problems with reasoning LLM, 2025. arXiv preprint arXiv:2503.10009

    work page arXiv 2025

  38. [38]

    A survey on large language models driven meta-optimizers for au- tomated intelligent optimization.Artificial Intelligence Review, 59(2):72, 2026

    Yan Zheng, Lida Zhang, Kaiwen Li, Rui Wang, Wenhua Li, Tao Zhang, Qingfu Zhang, and Yaochu Jin. A survey on large language models driven meta-optimizers for au- tomated intelligent optimization.Artificial Intelligence Review, 59(2):72, 2026. doi: 10.1007/s10462-025-11470-w

    work page doi:10.1007/s10462-025-11470-w 2026

  39. [39]

    Large Language Models as Optimization Controllers: Adaptive Continuation for SIMP Topology Optimization

    Shaoliang Yang, Jun Wang, and Yunsheng Wang. Large language models as optimization controllers: Adaptive continuation for SIMP topology optimization. arXiv preprint, 2026. arXiv:2603.25099. 43

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  40. [40]

    AutoSiMP: Autonomous topology opti- mization from natural language via LLM-driven problem configuration and adaptive solver control

    Shaoliang Yang, Jun Wang, and Yunsheng Wang. AutoSiMP: Autonomous topology opti- mization from natural language via LLM-driven problem configuration and adaptive solver control. arXiv preprint, 2026. arXiv:2603.27000

    work page arXiv 2026

  41. [41]

    Gemini 3.1 flash-lite: Built for intelligence at scale

    The Gemini Team. Gemini 3.1 flash-lite: Built for intelligence at scale. https://blog.google/innovation-and-ai/models-and-research/gemini-models/ gemini-3-1-flash-lite, March 2026. Official Google model announcement; accessed 2026-04-28

    work page 2026

  42. [42]

    Stress constrained multi- material topology optimization with the ordered SIMP method.Computer Methods in Applied Mechanics and Engineering, 373:113453, 2021

    Shuzhi Xu, Jikai Liu, Bin Zou, Quhao Li, and Yongsheng Ma. Stress constrained multi- material topology optimization with the ordered SIMP method.Computer Methods in Applied Mechanics and Engineering, 373:113453, 2021. doi: 10.1016/j.cma.2020.113453

    work page doi:10.1016/j.cma.2020.113453 2021

  43. [43]

    Stress-constrained topology optimization of structures subjected to nonproportional loading.International Journal for Numerical Methods in Engineering, 124(12):2818–2836, 2023

    Gustav Granlund, Mathias Wallin, Daniel Tortorelli, and Seth Watts. Stress-constrained topology optimization of structures subjected to nonproportional loading.International Journal for Numerical Methods in Engineering, 124(12):2818–2836, 2023. doi: 10.1002/ nme.7230

    work page 2023

  44. [44]

    Fail-safe stress-constrained manufactur- ing error tolerant topology optimization.Computer Methods in Applied Mechanics and Engineering, 421:116817, 2024

    Gustavo Assis da Silva and Hélio Emmendoerfer. Fail-safe stress-constrained manufactur- ing error tolerant topology optimization.Computer Methods in Applied Mechanics and Engineering, 421:116817, 2024. doi: 10.1016/j.cma.2024.116817

    work page doi:10.1016/j.cma.2024.116817 2024

  45. [45]

    Minimum stress optimal design with the level set method.Engineering Analysis with Boundary Elements, 32(11):909–918, 2008

    Grégoire Allaire and François Jouve. Minimum stress optimal design with the level set method.Engineering Analysis with Boundary Elements, 32(11):909–918, 2008. doi: 10. 1016/j.enganabound.2007.05.007

    work page 2008

  46. [46]

    Stress-related topology opti- mization via level set approach.Computer Methods in Applied Mechanics and Engineering, 200(47–48):3439–3452, 2011

    Xu Guo, Wei Sheng Zhang, Michael Yu Wang, and Peng Wei. Stress-related topology opti- mization via level set approach.Computer Methods in Applied Mechanics and Engineering, 200(47–48):3439–3452, 2011. doi: 10.1016/j.cma.2011.08.016

    work page doi:10.1016/j.cma.2011.08.016 2011

  47. [47]

    Alicia Kim

    Renato Picelli, Scott Townsend, Christopher Brampton, Julian Norato, and H. Alicia Kim. Stress-basedshapeandtopologyoptimizationwiththelevelsetmethod.Computer Methods in Applied Mechanics and Engineering, 329:1–23, 2018. doi: 10.1016/j.cma.2017.09.001

    work page doi:10.1016/j.cma.2017.09.001 2018

  48. [48]

    Fancello

    Hélio Emmendoerfer and Eduardo A. Fancello. Topology optimization with local stress constraint based on level set evolution via reaction–diffusion.Computer Methods in Applied Mechanics and Engineering, 305:62–88, 2016. doi: 10.1016/j.cma.2016.02.024

    work page doi:10.1016/j.cma.2016.02.024 2016

  49. [49]

    Stress-based topology optimization using bi-directional evolutionary structural optimization method.Computer Methods in Applied Mechanics and Engineering, 333:356–370, 2018

    Liang Xia, Lin Zhang, Qi Xia, and Tielin Shi. Stress-based topology optimization using bi-directional evolutionary structural optimization method.Computer Methods in Applied Mechanics and Engineering, 333:356–370, 2018. doi: 10.1016/j.cma.2018.01.035. 44

    work page doi:10.1016/j.cma.2018.01.035 2018