Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Graph Neural Network-Based Topology Optimization for Self-Supporting Structures in Additive Manufacturing

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A graph neural network optimizes parts to be self-supporting for 3D printing and globally stress-limited in one backpropagation loop.

desk verdict A credible AD-based integration of the Langelaar AM filter and p-norm stress into GNN topology optimization, but the stress-constrained claim is only validated against the same p-norm being penalized, not against actual peak stress. read the letter →

arxiv 2508.19169 v1 pith:GVSNJYBQ submitted 2025-08-26 cs.CE

Alireza Tabarraei , Saquib Ahmad Bhuiyan This is my paper
classification cs.CE
keywords topologyoptimizationgraphneuralnetworkadditivemanufacturingself-supportingstructuresoverhangconstraintstressvonMisesautomaticdifferentiation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that a graph neural network can serve as the optimizer itself, not just a surrogate, for topology optimization under additive-manufacturing constraints. The network predicts a continuous material density over the finite element mesh; a differentiable layer-by-layer overhang filter then enforces self-supporting printability, and a differentiable p-norm of von Mises stress enforces a global strength limit. Because every module sits in one differentiable computational graph, gradients flow back through the filter and stress aggregation automatically, with no hand-derived adjoint sensitivities. On benchmark beams and cantilevers, the framework produces topologies that are printable and stress-constrained, at the cost of a modest rise in compliance. If the approach holds, it shortens the path from a structural design problem to a ready-to-print, mechanically reliable part.

What carries the argument

The design variables are the weights of a Chebyshev spectral graph convolution network that acts as a neural field over the mesh: elements are nodes, and Fourier feature maps encode their spatial positions. The printed density is produced by a differentiable layer-by-layer overhang filter (smooth approximations of the min/max recurrence that propagates support upward), and mechanical response comes from a differentiable finite element solve. The global stress measure is a p-norm aggregation of element von Mises stresses scaled by sqrt(E), which is what the optimizer actually controls. Automatic differentiation through filter, solver, and stress aggregation is the mechanism that removes the n

What would settle it

Recompute the finite element stress field of the reported final designs (e.g., beam compliance 35.65e3, allowable stress 2.3) and compare the true maximum von Mises stress, computed on a refined mesh, against the allowable stress. If the true maximum exceeds 2.3 while the p-norm satisfies the constraint, the paper's stress-compliance claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim: a single differentiable graph can carry out the entire topology-optimization loop for printed parts. The mesh becomes a graph whose nodes are elements with Fourier-encoded coordinates; a Chebyshev graph convolutional network outputs a pseudo-density per element. A differentiable layer-by-layer overhang filter maps these to printable densities, and element von Mises stresses are aggregated into a scaled p-norm global stress constraint. Minimizing a loss of compliance, volume, and squared stress violation by backpropagation through the network weights replaces manual adjoint sensitivity derivations. Benchmarks (simply supported beam and two cantilevers) yield self-su

Load-bearing premise

The load-bearing assumption is that the smoothed, aggregated stress number the optimizer controls tracks the true peak stress closely enough; the paper never checks the actual largest element stress in the final designs.

Editorial extensions

If this is right

  • Topology optimization with AM and stress constraints no longer requires hand-derived adjoint sensitivities; any differentiable constraint can be appended to the loss and optimized by backpropagation.
  • The reported compliance increases (30.00e3 to 35.65e3 for the beam; roughly 177.9 to 192.8 for cantilevers) quantify the price of printability and stress control in stiffness terms.
  • Because the density field comes from a graph defined on the mesh, the same machinery applies to non-uniform and 3D meshes, where manual sensitivity recursions become unwieldy.
  • Designs meeting the layer-by-layer overhang criterion can be produced without sacrificial supports, reducing material waste and post-processing.
  • The continuous neural density output may act as an implicit regularization that smooths the topology and suppresses checkerboard patterns.

Reading between the lines

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

  • The stress-compliance claim is only as strong as the p-norm proxy; the paper never reports the actual maximum element stress of the final designs, so a verification step (or a corrected p-norm such as KS aggregation) is the immediate test.
  • The pipeline optimizes each problem from scratch; its efficiency claim is about skipping manual sensitivity derivations, not about amortizing learning across many designs, and transfer across load cases remains untested.
  • The smoothness of the neural field likely acts as an implicit density filter; isolating this effect would clarify whether the GNN replaces standard filtering steps or merely hides them.
  • The mechanical-reliability language should be read as stiffness plus simple von Mises proxy, not fatigue or fracture guarantees; production validation would demand material-specific failure models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a fully differentiable, graph neural network (GNN) based topology optimization framework for additive manufacturing. The design domain is discretized as a graph, a Chebyshev GNN with Fourier feature inputs predicts a blueprint density field, and a layer-by-layer differentiable AM filter based on Langelaar's overhang filter converts it to a printable density. The forward pass solves linear elasticity, computes compliance and a p-norm aggregated von Mises stress, and the loss (Eq. 24) combines compliance, volume, and stress penalty terms. Optimization is performed end-to-end with automatic differentiation over the GNN weights, avoiding manual adjoint derivations. Three benchmark problems (simply supported beam, tip-loaded cantilever, mid-loaded cantilever) are presented, showing qualitative differences between unconstrained, AM-filtered, and AM+stress-constrained designs, with compliance increases (e.g., 30.00e3 to 35.65e3 for the simply supported beam) as constraints are added.

Significance. If the stress-constrained and self-supporting claims are validated, the framework is a useful demonstration of a fully differentiable, sensitivity-free topology optimization pipeline that couples manufacturing constraints with mechanical reliability. The paper's strengths are its end-to-end differentiable architecture, use of a standard and physically motivated AM filter, and explicit treatment of stress via a smooth p-norm aggregation. The qualitative topologies (inclined members, material redistribution near stress concentrations) are plausible and consistent with prior work. However, the paper ships no code, provides no machine-checked proofs, and the validation of the central stress constraint is self-referential: the only reported stress metric is the penalized quantity in the loss. The self-supporting claim is likewise supported only by visual inspection. These gaps are fixable and do not invalidate the approach, but they currently leave the paper's headline conclusions under-supported.

major comments (3)
  1. [Section 8.1, Eqs. (16)-(17), (24)] The claim that the global stress constraint is satisfied is not supported by the evidence. The p-norm plotted in Figs. 6, 10, and 14 is exactly the term penalized in Eq. (24), using stresses scaled by sqrt(E_e) and averaged over all N elements. This is not the physical maximum von Mises stress: the sqrt(E_e) factor suppresses low-density elements, and a finite-p norm underestimates max_e sigma_vm unless p is large or a correction factor is used. The value of p is never stated, and the actual maximum element von Mises stress in any final design is never reported. I ask the authors to report the unscaled maximum (or 99th percentile) von Mises stress for each final design, state the p used, and, if desired, apply a p-norm correction such as in Ref. [40] and report both corrected and uncorrected values. Without this, the mechanical-reliability conclusion in Section 9 is not verifiable.
  2. [Section 3, Eq. (4); Figs. 5, 9, 13] The claim that the designs are 'self-supporting' is supported only by visual inspection of inclined members. The smooth AM filter in Eq. (4) uses finite P and epsilon, which only approximate the exact min/max operations in Eq. (3); no values for P and epsilon are reported. The paper does not quantify the maximum overhang angle, the number of unsupported elements, or any post-optimization check against the nominally 45-degree criterion. Since the central contribution is manufacturability, a quantitative verification (e.g., a geometric overhang-angle sweep of the final density field, or reporting the residual violation of the exact filter) is necessary.
  3. [Section 8, overall experiments] The paper does not report most hyperparameters needed to reproduce the experiments or to assess the strength of the stress penalty: the p in Eqs. (17)/(24), penalty weights alpha and gamma, P and epsilon in Eq. (4), the Fourier frequency matrix B (dimension m, sampling distribution), number of GNN layers/hidden units, learning rate, iteration count, and J0 normalization. Only the Chebyshev degree K=1 and the allowable stress value are given. The code is only available 'upon reasonable request', not as a stable artifact. These omissions are particularly problematic for the stress constraint, since the choice of p directly determines how tightly the p-norm bounds the maximum stress. Please add a table of all numerical settings.
minor comments (5)
  1. [Section 7, Eq. (33)] The equation numbering is inconsistent: Eq. (23) is followed by Eq. (33) for the Fourier feature mapping, while the loss in the same section is numbered Eq. (24). This suggests missing or mislabeled equations; please renumber.
  2. [Section 8.3, caption of Fig. 13(c)] The reported compliance for the mid-loaded cantilever with AM+stress constraint is 192.83, identical to the tip-loaded cantilever case in Fig. 9(c). This is either a coincidence or a copy-paste error; please verify and correct.
  3. [Section 8.3, text] The text says the evolution for the mid-loaded cantilever is 'presented in Fig. 15', but Fig. 14 contains the evolution curves and Fig. 15 contains the compliance comparison. Please correct the cross-reference.
  4. [Abstract and Section 9] The abstract and conclusion state that a 'global stress constraint is implemented' and 'enforced', but Eq. (24) uses a quadratic penalty, not a hard constraint. The paper should distinguish between a penalty formulation and a rigorously enforced constraint, especially if final p-norm values are only near the allowable threshold.
  5. [Section 4.2, Eq. (18)] The analytic derivative in Eq. (18) appears to be missing the 1/N factor and the 1/p prefactor that follow from differentiating Eq. (17). Although the paper states these sensitivities are not needed in the proposed framework, the equation should still be correct as presented.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: stress-compliance claim is a proxy-validation gap, not a self-referential reduction; self-citations are not load-bearing.

full rationale

The paper's derivation chain is self-contained. The GNN outputs blueprint densities, which pass through the AM filter and FEA solver; compliance, volume, and stress are computed from the simulated response, and the loss in Eq. 24 directly penalizes those quantities. There is no fitted parameter that is later renamed as a prediction. The stress constraint is a standard p-norm aggregation (Eq. 17) with independent prior support (Refs. [38-43]); the AM filter is from Langelaar (Ref. [18]), an external source. Self-citations (Refs. [30], [31], [44]) provide background and are not the sole justification for any load-bearing premise. The convergence plots (Figs. 6, 10, 14) show the same σPN that appears in the loss (Eq. 24) declining, which is a consistency check inherent to penalty methods, not circular evidence. However, the paper asserts that 'σPN≤0 ensures that the structure remains within the permissible stress range' (Sec. 4.1) without verifying against the actual maximum von Mises stress; finite-p p-norm with √E scaling can underestimate the true peak. This is a correctness/validation gap, not a circular derivation, so the circularity score is 0.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

Every adjustable knob needed to reproduce the experiments is counted above; none is derived from an external benchmark, and most are not even reported in the manuscript. The inventory is long partly because this is an engineering demonstration with hand-set constants (allowable stress 2.3, penalty weights, smoothness parameters) and partly because key SIMP and optimizer settings are omitted. No new physical entities are introduced; the 'neural field' is just the GNN parametrization of the density field (Section 7).

free parameters (7)
  • Allowable stress sigma_bar = 2.3 (given)
    Hand-set stress threshold in Eq. 24; the basis for the value is not justified, and all stress-constrained runs use it.
  • SIMP penalization exponent p and E_min = not reported
    Eq. 1 defines the material interpolation; the values needed to assemble K(rho) and to interpret the sqrt(Ee) stress scaling of Eq. 16 are never given.
  • Stress p-norm exponent p = not reported
    Eq. 17's aggregation exponent controls how closely sigma_PN approximates the maximum stress; not reported, so the tightness of the stress constraint is unknown.
  • Penalty weights alpha, gamma = not reported
    Eq. 24 weights the volume and stress penalty terms; the paper calls them 'constraint penalty parameters' without values, so the achieved trade-off is unreproducible.
  • AM filter smoothness parameters P and epsilon = epsilon_0 = 0.5 stated; P, epsilon not reported
    Eqs. 4a-c approximate the min/max filter; precision P and epsilon determine how close the printed density is to the exact Langelaar filter and therefore how genuinely self-supporting the result is.
  • Fourier feature frequency matrix B = not reported
    Eq. 33 uses 'a matrix of random frequencies'; the distribution and scale are not given, though they control the spatial resolution of the neural field.
  • Optimizer settings (learning rate, iterations) = not reported
    Adam is used; learning rate, schedule, and iteration counts appear in figures up to roughly 1500 but are never stated.
assumptions (6)
  • domain assumption SIMP material interpolation (Eq. 1) with penalization yields meaningful 0/1 designs
    Standard TO premise inherited from [34-37]; not re-derived in this paper.
  • domain assumption The Langelaar layerwise min/max overhang model (Eq. 3) accurately represents printability on regular meshes
    The self-supporting claim rests on this proxy from [18,19]; the smooth approximation (Eq. 4) with finite P and epsilon adds a further approximation that is not checked.
  • domain assumption Plane stress linear elasticity governs the structural response
    Section 4.1 states stresses are computed 'under plane stress conditions'; all conclusions are for 2D plane stress, not 3D printing reality.
  • domain assumption sqrt(Ee) scaling of von Mises stress (Eq. 16) yields a well-posed stress constraint
    Standard relaxation device against the SIMP stress singularity; it changes the controlled quantity from physical stress to a stiffness-weighted proxy, so 'stress-compliant' refers to the proxy.
  • ad hoc to paper A K=1 Chebyshev GNN with Fourier input features can represent near-optimal density fields
    K=1 means first-order neighborhood message passing (Section 8); expressiveness sufficient for the tested problems is assumed, not analyzed.
  • ad hoc to paper Adam on the nonconvex physics loss converges to practically good designs
    All results are single runs of one optimizer trajectory; no seeds, restarts, or convergence guarantees; the sharp early fluctuations in Figs. 6/10/14 show sensitivity to initialization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph Neural Network-Based Topology Optimization for Self-Supporting Structures in Additive Manufacturing." pith.science (2026). https://pith.science/paper/GVSNJYBQ

@misc{pith2026250819169,
  author       = {Pith},
  title        = {Pith review of: Graph Neural Network-Based Topology Optimization for Self-Supporting Structures in Additive Manufacturing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GVSNJYBQ}},
  note         = {Machine review of arXiv:2508.19169}
}
read the original abstract

This paper presents a machine learning-based framework for topology optimization of self-supporting structures, specifically tailored for additive manufacturing (AM). By employing a graph neural network (GNN) that acts as a neural field over the finite element mesh, the framework effectively learns and predicts continuous material distributions. An integrated AM filter ensures printability by eliminating unsupported overhangs, while the optimization process minimizes structural compliance under volume and stress constraints. The stress constraint is enforced using a differentiable p-norm aggregation of von Mises stress, promoting mechanical reliability in the optimized designs. A key advantage of the approach lies in its fully differentiable architecture, which leverages automatic differentiation throughout the optimization loop--eliminating the need for explicit sensitivity derivation for both the filter and the stress constraint. Numerical experiments demonstrate the ability of the framework to generate stress-constrained manufacturable topologies under various loading and boundary conditions, offering a practical pathway toward AM-ready high-performance designs with reduced post-processing requirements.

Figures

Figures reproduced from arXiv: 2508.19169 by the authors.

Figure 1
Figure 1. Schematic diagram of AM filter most common objective functions used in this context is the minimization of the compliance function, which is closely related to maximizing the stiffness of the struc￾ture. A volume constraint is often imposed to limit the amount of material used, corresponding to control￾ling the structure’s weight. The standard form of the topology optimization problem with these objectives and const… view at source ↗
Figure 2
Figure 2. Illustration of Proposed Framework [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Overview of the graph neural network architecture used for topology optimization. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: A simply supported beam under distributed loading. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Designed topology for a simply supported beam under a distributed load a) without AM filter, b) with AM filter c) with AM filter [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Evolution of compliance, P-Norm volume fraction over iterations for simply supported under a distributed load . [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 8
Figure 8. Figure 8: Tip Cantilever. and steadily decreases, reflecting the optimizer’s effort to minimize material usage while adhering to volume constraints. The dotted curve represents the p-norm of the von Mises stress, demonstrating that the proposed method successfully satisfies the …
Figure 9
Figure 9. Figure 9: Designed topology for tip cantilever. a) without AM filter, b) with AM filter, c) with AM filter [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Evolution of compliance, P-Norm volume fraction over iterations for tip Cantilever Beam . [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: Comparison of compliance all three conditions for the tip [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]
Figure 12
Figure 12. Figure 12: Mid Cantilever. The figure includes snapshots of the evolving topol￾ogy at different iteration stages. These images illus￾trate the progressive refinement of the structure: from an underdeveloped geometry in the early iterations to a well-defined and efficient design …
Figure 10
Figure 10. Figure 10: Initially, the compliance curve exhibits a high [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 13
Figure 13. Figure 13: Designed topology for mid cantilever. a) without AM filter, b) with AM filter, c) with AM filter [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]
Figure 14
Figure 14. Figure 14: Evolution of compliance, P-Norm volume fraction over iterations for mid Cantilever Beam . [PITH_FULL_IMAGE:figures/full_fig_p012_14.png]
Figure 15
Figure 15. Figure 15: Comparison of compliance all three conditions for the mid [PITH_FULL_IMAGE:figures/full_fig_p012_15.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Physics-Informed Transformer for Real-Time High-Fidelity Topology Optimization

    cs.CE 2026-04 unverdicted novelty 6.0 of 10

    A transformer model with self-attention and auxiliary physics losses learns a direct non-iterative mapping from loads and fields to manufacturable optimized topologies.

Reference graph

Works this paper leans on

47 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [40]

    K. Lee, K. Ahn, J. Yoo, A novel p-norm correction method for lightweight topology optimization under maximum stress con- straints, Computers & Structures 171 (2016) 18–30

    work page 2016

  2. [1]

    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 manufac- turing, Chinese Journal of Aeronautics 33 (4) (2020) 1252– 1259

    work page 2020

  3. [2]

    Zhu, W.-H

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

    work page 2016

  4. [3]

    S. S. Hanush, M. Manjaiah, Topology optimization of aerospace part to enhance the performance by additive manufacturing pro- cess, Materials Today: Proceedings 62 (2022) 7373–7378

    work page 2022

  5. [4]

    M. K. Leader, T. W. Chin, G. Kennedy, High resolution topology optimization of aerospace structures with stress and frequency constraints, in: 2018 Multidisciplinary Analysis and Optimiza- tion Conference, 2018, p. 4056

    work page 2018

  6. [5]

    J. Rao, S. Kiran, J. Kamesh, M. A. Padmanabhan, S. Chandra, Topology optimization of aircraft wing, Journal of Aerospace Sciences and Technologies (2009) 402–414

    work page 2009

  7. [6]

    Jankovics, A

    D. Jankovics, A. Barari, Customization of automotive structural components using additive manufacturing and topology opti- mization, IFAC-PapersOnLine 52 (10) (2019) 212–217

    work page 2019

  8. [7]

    Cavazzuti, A

    M. Cavazzuti, A. Baldini, E. Bertocchi, D. Costi, E. Torricelli, P. Moruzzi, High performance automotive chassis design: a topology optimization based approach, Structural and Multidis- ciplinary Optimization 44 (2011) 45–56

    work page 2011

Show all 47 references
  1. [8]

    M. N. Sudin, M. M. Tahir, F. R. Ramli, S. A. Shamsuddin, Topology optimization in automotive brake pedal redesign, In- ternational Journal of Engineering and Technology (IJET) 6 (1) (2014) 398–402

    2014

  2. [9]

    A. R. Yıldız, U. A. Kılıc ¸arpa, E. Demirci, M. Do ˘gan, Topogra- phy and topology optimization of diesel engine components for light-weight design in the automotive industry, Materials Test- ing 61 (1) (2019) 27–34

    2019

  3. [10]

    Tyflopoulos, Topology optimization in mechanical engineer- ing: Implementation and practical aspects (2022)

    E. Tyflopoulos, Topology optimization in mechanical engineer- ing: Implementation and practical aspects (2022). 13

    2022

  4. [11]

    Leary, L

    M. Leary, L. Merli, F. Torti, M. Mazur, M. Brandt, Optimal topology for additive manufacture: A method for enabling ad- ditive manufacture of support-free optimal structures, Materials & Design 63 (2014) 678–690

    2014

  5. [12]

    A. T. Gaynor, J. K. Guest, Topology optimization considering overhang constraints: Eliminating sacrificial support material in additive manufacturing through design, Structural and Multidis- ciplinary Optimization 54 (5) (2016) 1157–1172

    2016

  6. [13]

    T. E. Johnson, A. T. Gaynor, Three-dimensional projection- based topology optimization for prescribed-angle self- supporting additively manufactured structures, Additive Manufacturing 24 (2018) 667–686

    2018

  7. [14]

    X. Qian, Undercut and overhang angle control in topology op- timization: a density gradient based integral approach, Interna- tional Journal for Numerical Methods in Engineering 111 (3) (2017) 247–272

    2017

  8. [15]

    C. Wang, X. Qian, Simultaneous optimization of build orienta- tion and topology for additive manufacturing, Additive Manu- facturing 34 (2020) 101246

    2020

  9. [16]

    Kuo, C.-C

    Y .-H. Kuo, C.-C. Cheng, Y .-S. Lin, C.-H. San, Support struc- ture design in additive manufacturing based on topology opti- mization, Structural and Multidisciplinary Optimization 57 (1) (2018) 183–195

    2018

  10. [17]

    A. M. Mirzendehdel, K. Suresh, Support structure constrained topology optimization for additive manufacturing, Computer- Aided Design 81 (2016) 1–13

    2016

  11. [18]

    Langelaar, Topology optimization of 3D self-supporting structures for additive manufacturing, Additive manufacturing 12 (2016) 60–70

    M. Langelaar, Topology optimization of 3D self-supporting structures for additive manufacturing, Additive manufacturing 12 (2016) 60–70

    2016

  12. [19]

    Langelaar, An additive manufacturing filter for topology optimization of print-ready designs, Structural and multidisci- plinary optimization 55 (2017) 871–883

    M. Langelaar, An additive manufacturing filter for topology optimization of print-ready designs, Structural and multidisci- plinary optimization 55 (2017) 871–883

    2017

  13. [20]

    M. Langelaar, Combined optimization of part topology, support structure layout and build orientation for additive manufactur- ing, Structural and Multidisciplinary Optimization 57 (5) (2018) 1985–2004

    2018

  14. [21]

    van de Ven, R

    E. van de Ven, R. Maas, C. Ayas, M. Langelaar, F. van Keulen, Overhang control in topology optimization: a comparison of continuous front propagation-based and discrete layer-by-layer overhang control, Structural and Multidisciplinary Optimization 64 (2021) 761–778

    2021

  15. [22]

    Mart ´ınez-Frutos, P

    J. Mart ´ınez-Frutos, P. J. Mart ´ınez-Castej´on, D. Herrero-P ´erez, Efficient topology optimization using gpu computing with mul- tilevel granularity, Advances in Engineering Software 106 (2017) 47–62

    2017

  16. [23]

    N. Aage, E. Andreassen, B. S. Lazarov, Topology optimization using petsc: An easy-to-use, fully parallel, open source topol- ogy optimization framework, Structural and Multidisciplinary Optimization 51 (2015) 565–572

    2015

  17. [24]

    C. M. Parrott, D. W. Abueidda, K. A. James, Multidisciplinary topology optimization using generative adversarial networks for physics-based design enhancement, Journal of Mechanical De- sign 145 (6) (2023) 061704

    2023

  18. [25]

    Maz ´e, F

    F. Maz ´e, F. Ahmed, Di ffusion models beat GANs on topology optimization, in: Proceedings of the AAAI conference on artifi- cial intelligence, V ol. 37, 2023, pp. 9108–9116

    2023

  19. [26]

    Zhang, G

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

    2025

  20. [27]

    Lutheran, S

    A. Lutheran, S. Das, A. Tabarraei, Latent space di ffusion for topology optimization, arXiv preprint arXiv:2508.05624 (2025)

    2025 arXiv

  21. [28]

    H. Chi, Y . Zhang, T. L. E. Tang, L. Mirabella, L. Dalloro, L. Song, G. H. Paulino, Universal machine learning for topol- ogy optimization, Computer Methods in Applied Mechanics and Engineering 375 (2021) 112739

    2021

  22. [29]

    Chandrasekhar, K

    A. Chandrasekhar, K. Suresh, TOuNN: topology opti- mization using neural networks, Structural and Mul- tidisciplinary Optimization 63 (3) (2021) 1135–1149. doi:https://doi.org/10.1007/s00158-020-02748-4

    2021 doi

  23. [30]

    M. I. R. Shishir, A. Tabarraei, Topology optimization using neu- ral network for stress constrained problems, in: ASME Interna- tional Mechanical Engineering Congress and Exposition, V ol. 87684, American Society of Mechanical Engineers, 2023, p. V011T12A006

    2023

  24. [31]

    Tabarraei, Variational quantum latent encoding for topology optimization, arXiv preprint arXiv:2506.17487 (2025)

    A. Tabarraei, Variational quantum latent encoding for topology optimization, arXiv preprint arXiv:2506.17487 (2025)

    2025 arXiv

  25. [32]

    Chandrasekhar, K

    A. Chandrasekhar, K. Suresh, Multi-material topology opti- mization using neural networks, Computer-Aided Design 136 (2021) 103017. doi:https://doi.org/10.1016/j.cad.2021.103017

    2021

  26. [33]

    M. I. R. Shishir, A. Tabarraei, Multi–materials topology op- timization using deep neural network for coupled thermo– mechanical problems, Computers & Structures 291 (2024) 107218

    2024

  27. [34]

    M. P. Bendsøe, Optimal shape design as a material distribu- tion problem, Structural optimization 1 (4) (1989) 193–202. doi:https://doi.org/10.1007/BF01650949

    1989 doi

  28. [35]

    Sigmund, Morphology-based black and white fil- ters for topology optimization, Structural and Mul- tidisciplinary Optimization 33 (4) (2007) 401–424

    O. Sigmund, Morphology-based black and white fil- ters for topology optimization, Structural and Mul- tidisciplinary Optimization 33 (4) (2007) 401–424. doi:https://doi.org/10.1007/s00158-006-0087-x

    2007 doi

  29. [36]

    Andreassen, A

    E. Andreassen, A. Clausen, M. Schevenels, B. S. Lazarov, O. Sigmund, E fficient topology optimization in matlab using 88 lines of code, Structural and Multidisciplinary Optimization 43 (1) (2011) 1–16. doi:https: //doi.org/10.1007/s00158-010- 0594-7

    2011 doi

  30. [37]

    M. Zhou, G. Rozvany, The COC algorithm, part ii: Topologi- cal, geometrical and generalized shape optimization, Computer methods in applied mechanics and engineering 89 (1-3) (1991) 309–336. doi:https://doi.org/10.1016/0045-7825(91)90046-9

    1991 doi

  31. [38]

    C. Y . Kiyono, S. L. Vatanabe, E. C. N. Silva, J. Reddy, A new multi-p-norm formulation approach for stress-based topology optimization design, Composite Structures 156 (2016) 10–19

    2016

  32. [39]

    Holmberg, B

    E. Holmberg, B. Torstenfelt, A. Klarbring, Stress constrained topology optimization, Structural and Multidisciplinary Opti- mization 48 (2013) 33–47

    2013

  33. [41]

    Z. Fan, L. Xia, W. Lai, Q. Xia, T. Shi, Evolutionary topology op- timization of continuum structures with stress constraints, Struc- tural and Multidisciplinary Optimization 59 (2019) 647–658

    2019

  34. [42]

    Zheng, G

    J. Zheng, G. Zhang, C. Jiang, Stress-based topology optimiza- tion of thermoelastic structures considering self-support con- straints, Computer Methods in Applied Mechanics and Engi- neering 408 (2023) 115957

    2023

  35. [43]

    Yaghoobi, M

    A. Yaghoobi, M. Asghari, H. Babaei, Introducing material- specific stress constraints in multi-material topology optimiza- tion, Computers & Structures 313 (2025) 107756

    2025

  36. [44]

    S. A. Bhuiyan, A. Tabarraei, Graph neural network-based topol- ogy optimization for e fficient support structure design in addi- tive manufacturing (2025)

    2025

  37. [45]

    M. He, Z. Wei, J.-R. Wen, Convolutional neural networks on graphs with chebyshev approximation, revisited, Advances in neural information processing systems 35 (2022) 7264–7276

    2022

  38. [46]

    D. P. Kingma, J. Ba, Adam: A method for stochas- tic optimization, arXiv preprint arXiv:1412.6980 (2017). doi:https://doi.org/10.48550/arXiv.1412.6980

    2017 doi arXiv

  39. [47]

    Zhang, Improved adam optimizer for deep neural net- works, in: 2018 IEEE /ACM 26th International Sympo- sium on Quality of Service (IWQoS), IEEE, 2018, pp

    Z. Zhang, Improved adam optimizer for deep neural net- works, in: 2018 IEEE /ACM 26th International Sympo- sium on Quality of Service (IWQoS), IEEE, 2018, pp. 1–2. 14 doi:https://doi.org/10.1109/IWQoS.2018.8624183. 15

    2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.