pith. sign in

arxiv: 2606.19921 · v1 · pith:6BZ4BHCCnew · submitted 2026-06-18 · 💻 cs.AI

eCNNTO: A Highly Generalizable ConvNet for Accelerating Topology Optimization

Shengbiao Lu , Xiaodong Wei This is my paper

Pith reviewed 2026-06-26 17:35 UTC · model grok-4.3

classification 💻 cs.AI
keywords topology optimizationconvolutional neural networkdensity based methodsacceleration techniquesgeneralizationfinite elementstructural optimizationmachine learning
0
0 comments X

The pith

An element-based convolutional network accelerates topology optimization by predicting final densities from early stages across new problems.

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

The paper presents eCNNTO, a convolutional neural network that speeds up density-based topology optimization. Standard methods take many iterations of finite element analysis, which slows down especially for fine meshes. eCNNTO trains on final density values from a small number of examples to predict near-optimal densities early in new optimizations, using residual connections to keep neighboring elements consistent. This allows it to handle changes in boundary conditions, loads, shapes, mesh sizes, and even non-design areas without retraining. As a result, it cuts the needed iterations by as much as 90 percent in two dimensions and 97 percent in three dimensions.

Core claim

The central claim is that training an element-wise convolutional neural network with residual connections on the final-stage density histories of a limited set of topology optimization problems enables accurate prediction of near-optimal densities from early histories on previously unseen problems that differ in boundary conditions, loading, geometry, mesh resolution, and non-design domains, thereby skipping the majority of iterations.

What carries the argument

Element-based convolutional neural network with residual connections, trained on final rather than early density histories.

If this is right

  • Topology optimization becomes feasible for much finer meshes without a matching rise in total computation.
  • The same trained model works for three-dimensional problems as well as two-dimensional ones.
  • Structures with non-design domains can be optimized without special handling or extra data.
  • Overall design time decreases substantially for repeated or varied structural problems.

Where Pith is reading between the lines

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

  • The method might allow integration into design tools for near-instant feedback on structural layouts.
  • Generalization to varying mesh resolutions suggests it could support multi-resolution optimization workflows.
  • If extended, the approach may apply to other iterative density-field optimizations like those in fluid dynamics.
  • Validation on problems with time-dependent loads or material nonlinearity would test the limits of the current training strategy.

Load-bearing premise

The density patterns at the end of optimization from a few training problems contain enough information for a CNN to predict correct early-stage densities that lead to valid connected topologies in many different new optimization scenarios.

What would settle it

Apply the trained network to an optimization problem with a loading case and domain geometry outside the training distribution, run the reduced iterations, and verify whether the resulting structure matches the fully optimized one in terms of connectivity and mechanical performance.

Figures

Figures reproduced from arXiv: 2606.19921 by Shengbiao Lu, Xiaodong Wei.

Figure 1
Figure 1. Figure 1: Evolution of element densities with respect to the number of iterations. In the [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall workflow of eCNNTO. (a) Preparation of the training dataset, where [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Network architecture of eCNNTO. Convolutional Feature Extractor (CFE) con [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Problem setups for dataset construction in 2D, where [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Problem setups for dataset construction in 3D, where [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Three options to select features for training: Early Stage (green), Middle Stage [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 6
Figure 6. Figure 6: As will be shown, different choice of the feature leads to greatly varying performance in both training and testing. Note that DLTOP only considered the Early Stage strategy. Indeed, both the Middle and Final Stage strategies appear counter-intuitive. However, the Final Stage proves to over-perform the Early Stage in every aspect, such as the speedup for TO, structural connectivity, and the required data s… view at source ↗
Figure 7
Figure 7. Figure 7: Two out-of-distribution examples, where Lx = 2, Ly = 1, and Vf = 0.5 is the volume fraction. (a) A cantilever beam subjected to a concentrated force F = 1 at its lower-right corner, and (b) a simply supported beam under a uniformly distributed force q = 1. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of predicted structures between DLTOP and eCNNTO, where the [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of predicted structures between DLTOP and eCNNTO. (a) A beam [PITH_FULL_IMAGE:figures/full_fig_p026_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Problem settings of 2D test problems, where [PITH_FULL_IMAGE:figures/full_fig_p028_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Optimized structures in 2D using SIMP and eCNNTO: (a, b) Long beam, (c, [PITH_FULL_IMAGE:figures/full_fig_p031_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Problem settings of 3D test problems, where [PITH_FULL_IMAGE:figures/full_fig_p034_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Optimized structures in 3D using SIMP and eCNNTO: (a, b) Cantilever beam [PITH_FULL_IMAGE:figures/full_fig_p036_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Topology optimization with non-design domains. (a) A cantilever beam with [PITH_FULL_IMAGE:figures/full_fig_p041_14.png] view at source ↗
read the original abstract

This work proposes an element-based Convolutional Neural Network (CNN) to accelerate density-based Topology Optimization (TO), termed eCNNTO. TO generally undergoes a large number of iterations, where finite element analysis is performed in every iteration, leading to the efficiency bottleneck especially when dense meshes are used to achieve high-resolution designs. To address this limitation, eCNNTO is proposed to build upon Kallioras et al. (2020), where a Deep Belief Network (DBN) was trained for every element to predict its near-optimal density from its early history, thereby skipping the great majority of iterations and significantly accelerating the TO procedure. However, the method lacks spatial correlations among neighboring elements and may lead to disconnected features in the final structure. The proposed method employs CNN with residual connections to address this issue. On top of it, a novel training strategy is introduced to further enhance the optimization efficiency, where the training dataset consists of the final stage density histories rather than early ones. This change can also help reduce the required training data size. eCNNTO requires only a small dataset to train and yet it can be generalized to problems with largely different boundary conditions, loading cases, design domain geometries, mesh resolutions, as well as non-design domains. In the end, the generalization capabilities and efficiency of eCNNTO are demonstrated through a variety of examples in two and three dimensions, achieving up to 90% and 97% reduction of iterations, respectively.

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 / 2 minor

Summary. The paper proposes eCNNTO, an element-based CNN with residual connections to accelerate density-based topology optimization. Building on Kallioras et al. (2020), it trains the network to predict near-optimal densities from early-iteration histories but uses final-stage density histories for training to reduce data needs and improve generalization. The method claims to handle problems with substantially different boundary conditions, loads, domain geometries, mesh resolutions, and non-design domains while achieving up to 90% (2D) and 97% (3D) iteration reductions, addressing disconnected features in prior DBN approaches via spatial correlations.

Significance. If the generalization and validity claims hold with low compliance errors and mechanically valid topologies, the work would provide a practical, data-efficient acceleration technique for high-resolution TO that improves on element-wise DBN methods. The novel final-stage training strategy and residual CNN architecture are potentially reusable ideas for surrogate-based optimization.

major comments (2)
  1. [Abstract] Abstract: the central generalization claim (to 'largely different' BCs, loads, geometries, meshes, and non-design domains) is stated without any quantitative metrics on prediction error, compliance deviation, constraint violation rates, or measures of distribution shift between training and test problems. This absence makes the 90%/97% iteration-reduction claim unverifiable from the provided text and is load-bearing for the acceleration result.
  2. [Abstract] Abstract and §1 (implied from context): the training change to final-stage densities is presented as enabling reliable early-history prediction on unseen problems, yet no evidence is supplied that the learned mapping avoids disconnected or invalid topologies (the exact failure mode attributed to the prior DBN). A concrete test (e.g., topology validity or mechanical feasibility metrics on held-out cases) is required to substantiate transfer.
minor comments (2)
  1. Notation for the element-wise CNN input (early density history tensor) and output (predicted density) should be defined explicitly, including how non-design domains are masked or handled.
  2. The manuscript should report the exact number of training problems, their mesh resolutions, and the precise differences (e.g., quantitative distance metrics) between training and test instances to allow readers to assess the claimed generalization.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We agree that the abstract requires strengthening with quantitative support and will revise accordingly. Our point-by-point responses to the major comments are provided below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central generalization claim (to 'largely different' BCs, loads, geometries, meshes, and non-design domains) is stated without any quantitative metrics on prediction error, compliance deviation, constraint violation rates, or measures of distribution shift between training and test problems. This absence makes the 90%/97% iteration-reduction claim unverifiable from the provided text and is load-bearing for the acceleration result.

    Authors: We agree that the abstract would benefit from explicit quantitative metrics to make the generalization and acceleration claims self-contained. The body of the manuscript presents these details through the experimental results on held-out problems. In the revised version we will update the abstract to include specific values drawn from those experiments, such as ranges for compliance deviation and prediction error. revision: yes

  2. Referee: [Abstract] Abstract and §1 (implied from context): the training change to final-stage densities is presented as enabling reliable early-history prediction on unseen problems, yet no evidence is supplied that the learned mapping avoids disconnected or invalid topologies (the exact failure mode attributed to the prior DBN). A concrete test (e.g., topology validity or mechanical feasibility metrics on held-out cases) is required to substantiate transfer.

    Authors: The CNN architecture with residual connections is introduced precisely to capture spatial correlations and thereby mitigate the disconnected-feature issue of element-wise DBNs. The experimental section shows resulting topologies that remain connected and mechanically feasible across the tested distribution shifts. We acknowledge that the abstract itself does not quantify topology-validity rates; we will add a concise reference to such metrics (computed on the held-out cases) in the revised abstract. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external simulation data and independent testing

full rationale

The paper trains an element-based residual CNN on density histories generated by standard topology optimization simulations (building on but not self-citing the 2020 DBN baseline). The claimed generalization and iteration reductions are demonstrated via explicit experiments on held-out problems with altered boundary conditions, loads, geometries, and meshes. No equation equates a reported speed-up to a fitted constant, no prediction reduces to its own training input by construction, and the central premise does not rest on a load-bearing self-citation chain. The method is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that element density histories contain sufficient information to predict near-optimal values and that convolutional layers can enforce spatial consistency without additional constraints.

free parameters (1)
  • CNN hyperparameters (layers, filters, residual blocks)
    Chosen to fit the element-wise prediction task; exact values not stated in abstract.
axioms (1)
  • domain assumption Density history of an element and its spatial neighborhood is predictive of its near-optimal density value.
    Inherited from the 2020 DBN baseline and extended to CNN.

pith-pipeline@v0.9.1-grok · 5790 in / 1304 out tokens · 24843 ms · 2026-06-26T17:35:14.748476+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

38 extracted references · 37 canonical work pages · 2 internal anchors

  1. [1]

    Nature , volume =

    Giga-Voxel Computational Morphogenesis for Structural Design , author =. Nature , volume =. doi:10.1038/nature23911 , urldate =

    work page doi:10.1038/nature23911

  2. [2]

    Heliyon , volume =

    Topology optimization of coronary artery stent considering structural and hemodynamic parameters , author =. Heliyon , volume =. doi:10.1016/j.heliyon.2024.e39452 , urldate =

    work page doi:10.1016/j.heliyon.2024.e39452 2024

  3. [3]

    Structural and Multidisciplinary Optimization , volume = 43, number = 1, pages =

    Andreassen, Erik and Clausen, Anders and Schevenels, Mattias and Lazarov, Boyan S. and Sigmund, Ole , year = 2011, month = jan, journal =. Efficient topology optimization in. doi:10.1007/s00158-010-0594-7 , urldate =

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

  4. [4]

    3D Topology Optimization using Convolutional Neural Networks

    Banga, Saurabh and Gehani, Harsh and Bhilare, Sanket and Patel, Sagar and Kara, Levent , year = 2018, month = aug, number =. doi:10.48550/arXiv.1808.07440 , urldate =. 1808.07440 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1808.07440 2018

  5. [5]

    Journal of Mechanical Design , pages =

    Behzadi, Mohammad Mahdi and Ilie. Journal of Mechanical Design , volume =. doi:10.1115/1.4052757 , urldate =

    work page doi:10.1115/1.4052757

  6. [6]

    Structural Optimization , volume = 1, number = 4, pages =

    Optimal shape design as a material distribution problem , author =. Structural optimization , volume =. doi:10.1007/BF01650949 , urldate =

    work page doi:10.1007/bf01650949

  7. [7]

    2004 , publisher=

    Bends. Topology. doi:10.1007/978-3-662-05086-6 , urldate =

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

  8. [8]

    Structural and Multidisciplinary Optimization , publisher =

    Chandrasekhar, Aaditya and Suresh, Krishnan , year = 2021, month = mar, journal =. doi:10.1007/s00158-020-02748-4 , urldate =

    work page doi:10.1007/s00158-020-02748-4 2021

  9. [9]

    Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil , year = 2021, month = jun, number =. An. doi:10.48550/arXiv.2010.11929 , urldate =. 2010.11929 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2010.11929 2021

  10. [10]

    Guo, Xu and Zhang, Weisheng and Zhong, Wenliang , year = 2014, month = aug, journal =. Doing. doi:10.1115/1.4027609 , urldate =

    work page doi:10.1115/1.4027609 2014

  11. [11]

    He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian , year = 2016, pages =. Deep. Proceedings of the

    2016

  12. [12]

    Engineering Applications of Artificial Intelligence , volume =

    Dynamic Graph-Based Convergence Acceleration for Topology Optimization in Unstructured Meshes , author =. Engineering Applications of Artificial Intelligence , volume =. doi:10.1016/j.engappai.2024.107916 , urldate =

    work page doi:10.1016/j.engappai.2024.107916 2024

  13. [13]

    Joo, Younghwan and Yu, Yonggyun and Jang, In Gwun , year = 2021, journal =. Unit. doi:10.1109/ACCESS.2021.3125014 , urldate =

    work page doi:10.1109/access.2021.3125014 2021

  14. [14]

    Structural and Multidisciplinary Optimization , volume =

    Accelerated Topology Optimization by Means of Deep Learning , author =. Structural and Multidisciplinary Optimization , volume =. doi:10.1007/s00158-020-02545-z , urldate =

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

  15. [15]

    Heaviside Projection Based Topology Optimization by a

    Kawamoto, Atsushi and Matsumori, Tadayoshi and Yamasaki, Shintaro and Nomura, Tsuyoshi and Kondoh, Tsuguo and Nishiwaki, Shinji , year = 2011, month = jul, journal =. Heaviside Projection Based Topology Optimization by a. doi:10.1007/s00158-010-0562-2 , urldate =

    work page doi:10.1007/s00158-010-0562-2 2011

  16. [16]

    Nature521(7553), 436–444 (2015) https://doi.org/10.1038/nature14539

    Deep Learning , author =. Nature , volume =. doi:10.1038/nature14539 , urldate =

    work page doi:10.1038/nature14539

  17. [17]

    Engineering Applications of Artificial Intelligence , volume =

    Accelerating Topology Optimization Using Deep Learning-Based Image Super-Resolution , author =. Engineering Applications of Artificial Intelligence , volume =. doi:10.1016/j.engappai.2024.108370 , urldate =

    work page doi:10.1016/j.engappai.2024.108370 2024

  18. [18]

    International Journal of Heat and Mass Transfer , volume =

    Genetic Algorithm Based Topology Optimization of Heat Exchanger Fins Used in Aerospace Applications , author =. International Journal of Heat and Mass Transfer , volume =. doi:10.1016/j.ijheatmasstransfer.2021.121002 , urldate =

    work page doi:10.1016/j.ijheatmasstransfer.2021.121002 2021

  19. [19]

    doi:10.1115/1.4049533 , urldate =

    Nie, Zhenguo and Lin, Tong and Jiang, Haoliang and Kara, Levent Burak , year = 2021, month = feb, journal =. doi:10.1115/1.4049533 , urldate =

    work page doi:10.1115/1.4049533 2021

  20. [20]

    Engineering Applications of Artificial Intelligence , volume =

    Deep Learning Accelerated Efficient Framework for Topology Optimization , author =. Engineering Applications of Artificial Intelligence , volume =. doi:10.1016/j.engappai.2024.108559 , urldate =

    work page doi:10.1016/j.engappai.2024.108559 2024

  21. [21]

    Structural and Multidisciplinary Optimization , volume =

    Accelerating Gradient-Based Topology Optimization Design with Dual-Model Artificial Neural Networks , author =. Structural and Multidisciplinary Optimization , volume =. doi:10.1007/s00158-020-02770-6 , urldate =

    work page doi:10.1007/s00158-020-02770-6

  22. [22]

    Computer Methods in Applied Mechanics and Engineering , volume =

    Isogeometric Topology Optimization of Thin-Walled Structures with Complex Design Domains , author =. Computer Methods in Applied Mechanics and Engineering , volume =. doi:10.1016/j.cma.2025.118114 , urldate =

    work page doi:10.1016/j.cma.2025.118114 2025

  23. [23]

    , year = 2001, month = apr, journal =

    Sigmund, O. , year = 2001, month = apr, journal =. A 99 Line Topology Optimization Code Written in. doi:10.1007/s001580050176 , urldate =

    work page doi:10.1007/s001580050176 2001

  24. [24]

    Structural and Multidisciplinary Optimization , volume =

    Morphology-Based Black and White Filters for Topology Optimization , author =. Structural and Multidisciplinary Optimization , volume =. doi:10.1007/s00158-006-0087-x , urldate =

    work page doi:10.1007/s00158-006-0087-x

  25. [25]

    and Petersson, J

    Sigmund, O. and Petersson, J. , year = 1998, month = aug, journal =. Numerical Instabilities in Topology Optimization:. doi:10.1007/BF01214002 , urldate =

    work page doi:10.1007/bf01214002 1998

  26. [26]

    Structural and Multidisciplinary Optimization , volume =

    Sensitivity Filtering from a Continuum Mechanics Perspective , author =. Structural and Multidisciplinary Optimization , volume =. doi:10.1007/s00158-012-0814-4 , urldate =

    work page doi:10.1007/s00158-012-0814-4

  27. [27]

    Russian Journal of Numerical Analysis and Mathematical Modelling , volume =

    Neural Networks for Topology Optimization , author =. Russian Journal of Numerical Analysis and Mathematical Modelling , volume =. doi:10.1515/rnam-2019-0018 , urldate =

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

  28. [28]

    The International Journal of Advanced Manufacturing Technology , volume =

    Topology Optimization and Lightweight Design of Stamping Dies for Forming Automobile Panels , author =. The International Journal of Advanced Manufacturing Technology , volume =. doi:10.1007/s00170-022-09683-2 , urldate =

    work page doi:10.1007/s00170-022-09683-2

  29. [29]

    International Journal for Numerical Methods in Engineering , volume =

    The Method of Moving Asymptotes---a New Method for Structural Optimization , author =. International Journal for Numerical Methods in Engineering , volume =. doi:10.1002/nme.1620240207 , urldate =

    work page doi:10.1002/nme.1620240207

  30. [30]

    Efficient Large-Scale

    Wang, Junpeng and Aage, Niels and Wu, Jun and Sigmund, Ole and Westermann, R. Efficient Large-Scale. Structural and Multidisciplinary Optimization , volume =. doi:10.1007/s00158-025-04127-3 , urldate =

    work page doi:10.1007/s00158-025-04127-3

  31. [31]

    Computer Methods in Applied Mechanics and Engineering , volume =

    A Level Set Method for Structural Topology Optimization , author =. Computer Methods in Applied Mechanics and Engineering , volume =. doi:10.1016/S0045-7825(02)00559-5 , urldate =

    work page doi:10.1016/s0045-7825(02)00559-5

  32. [32]

    Applied Mathematical Modelling , volume =

    Phase-Field Method Combined with Optimality Criteria Approach for Topology Optimization , author =. Applied Mathematical Modelling , volume =. doi:10.1016/j.apm.2024.02.006 , urldate =

    work page doi:10.1016/j.apm.2024.02.006 2024

  33. [33]

    Structural and Multidisciplinary Optimization , volume =

    On the Use of Artificial Neural Networks in Topology Optimisation , author =. Structural and Multidisciplinary Optimization , volume =. doi:10.1007/s00158-022-03347-1 , urldate =

    work page doi:10.1007/s00158-022-03347-1

  34. [34]

    Bi-directional

    Xia, Liang and Xia, Qi and Huang, Xiaodong and Xie, Yi Min , year = 2018, month = apr, journal =. Bi-directional. doi:10.1007/s11831-016-9203-2 , urldate =

    work page doi:10.1007/s11831-016-9203-2 2018

  35. [35]

    Computers & Structures , volume =

    A Simple Evolutionary Procedure for Structural Optimization , author =. Computers & Structures , volume =. doi:10.1016/0045-7949(93)90035-C , urldate =

    work page doi:10.1016/0045-7949(93)90035-c

  36. [36]

    Thin-Walled Structures , volume =

    An Online Autonomous Learning and Prediction Scheme for Machine Learning Assisted Structural Optimization , author =. Thin-Walled Structures , volume =. doi:10.1016/j.tws.2022.110500 , urldate =

    work page doi:10.1016/j.tws.2022.110500 2022

  37. [37]

    Design of

    Xue, Huipeng and Luo, Zhen and Brown, Terry and Beier, Susann , year = 2020, month = jul, journal =. Design of. doi:10.3389/fbioe.2020.00736 , urldate =

    work page doi:10.3389/fbioe.2020.00736 2020

  38. [38]

    Structural and Multidisciplinary Optimization , volume =

    An Integrated Multi-Objective Topology Optimization Method for Automobile Wheels Made of Lightweight Materials , author =. Structural and Multidisciplinary Optimization , volume =. doi:10.1007/s00158-021-02913-3 , urldate =

    work page doi:10.1007/s00158-021-02913-3