TopoDiff: Diffusion-Model Topology Optimization and Its Evaluation Metrics

New
License: CC BY 4.0 for the arXiv version; published at AAAI 2023 · Updated 2026-11-16
My rating

1. Overview

Topology optimization is the technology that turns the question "where should material sit, given a design domain and a set of loads" into an optimisation problem. It has established uses in aerospace, automotive and civil engineering, and it has a clear commercial path: mainstream CAD tools ship generative design as a native capability, where a designer enters loads, space envelopes, materials and process constraints and the software returns several candidate geometries. The tool answers whether the problem can be computed. It does not answer how to make the resulting geometry actually usable.

Generative methods entered this field with an appealing speed argument and a new class of problems. Traditional SIMP-type methods iterate finite element analysis and sensitivity derivatives, re-solving the structure at every step, so cost rises with resolution. Deep generative models recast optimisation as a single forward pass, appearing to bypass the iterative loop entirely. But one sampling pass does not yield one usable structure. Pixel-level similarity does not imply mechanical similarity, because material placed in the wrong location can look similar while remaining weak; and a generated layout containing floating material can look clean while being impossible to manufacture. Both defects are silent and cannot be caught by looking at pictures.

TopoDiff matters because it turns these two issues into measurable quantities. The paper does not settle for pixel error: it defines separate geometric, physical and manufacturing metrics, and it uses a dual-level test protocol that separates generalisation to unseen constraint combinations within the training distribution from generalisation to out-of-distribution boundary conditions. That protocol is arguably more valuable than the method itself, since most subsequent generative topology-optimisation work builds on it.

There is also a practical reason to read this entry. The earlier description on this site presented TopoDiff as a conditional diffusion method from MIT, listed the claims of roughly eightfold lower performance error and elevenfold fewer infeasible samples as unconditional conclusions, and recorded the licence status as an arXiv preprint. Each needs correction. The paper was published in the AAAI 2023 proceedings and is no longer a preprint. Those two factors refer to the out-of-distribution level of the dual-level protocol, and on the in-distribution level the directions invert: performance error improves by roughly elevenfold while floating material improves by roughly 8.4-fold. More importantly, the earlier description said nothing about inference cost. The paper itself reports 0.06 seconds per generated topology for the conditional GAN and 21.59 seconds for TopoDiff, a difference of about 360 times. A method that replaces iterative optimisation with a single sampling pass is not faster than iterative optimisation when that single pass is slower than the iteration.

This entry is therefore rewritten as a fact-checked account: it restores the mechanism and the dataset composition, explains each metric as the paper defines it, marks the scope within which each factor claim holds, and folds in inference cost, the 3D extension and follow-up work. Everything was checked on 11 November 2026 against the arXiv full text, the AAAI publication record, the author project page and the follow-up paper.

2. How It Works

TopoDiff has three parts: a conditional diffusion generator, two surrogate models (a compliance regressor and a floating-material classifier), and a guidance mechanism that connects the surrogates to the denoising process.

The underlying task is structural topology optimisation: determine a material distribution inside a design domain so that an objective, here mechanical compliance, is minimised subject to volume fraction, displacement boundary conditions and load constraints. Traditional methods are built on SIMP, which assigns every mesh element a continuous density, pushes densities toward 0 or 1 via a penalisation factor and filtering, and then iterates using finite elements and gradients. Two weaknesses of that route are known: it is slow, and penalisation plus post-processing can yield suboptimal solutions.

The conditional diffusion generator is a DDPM-style discrete-time Gaussian diffusion model. The forward process adds noise to the structure according to a variance schedule, and the reverse process denoises step by step with a parameterised Gaussian, sampling from the prior and iterating down to the structure image. The denoiser is a UNet-based network with attention layers, and conditional information enters as extra channels. The paper cites the result that the covariance term may be fixed as a constant rather than learned.

The choice of conditioning representation is itself a design decision. Raw loads and boundary conditions are sparse, a handful of vectors on constrained nodes. The paper instead encodes conditions through physical fields, including volume fraction, von Mises stress and strain energy density, with the von Mises stress defined as the root of the squared deviatoric stress combination and the strain energy density as one half of the stress-strain contraction. Using physical fields instead of sparse inputs eases generalisation from sparse conditioning.

The guidance mechanism is the central innovation. The compliance regressor predicts the mechanical compliance of the current noisy topology, and regression guidance shifts the denoising mean toward low-compliance regions; the floating-material classifier outputs the probability that the structure contains no floating material, and classification guidance shifts the mean toward connected topologies. A first-order approximation of regression guidance gives a mean offset proportional to the gradient of the regressor, with a scale parameter controlling its strength, while classification guidance contributes the gradient of the log probability. When the two are combined, classification guidance is applied first and the compliance gradient is then evaluated at the already-shifted mean, which avoids both gradients being fixed at the original predicted mean. A maximum noise level threshold additionally gates each guidance term, since surrogate predictions are unreliable at very high noise.

The surrogate training data is likewise produced by physical simulation. The compliance regressor uses a UNet encoder adapted to regression, trained on roughly seventy-two thousand labelled samples drawn from the main dataset's optimal topologies, from non-optimal topologies generated by a fake-load procedure, and from the unconditional diffusion model's own outputs, with anomalous structures filtered by a compliance threshold. The fake-load procedure adds an extra load to the SIMP input constraints so that SIMP optimises for that extra load; the resulting structure is usually not optimal for the true load while still roughly satisfying the requirement that material exists where loads are applied. The floating-material classifier is trained on seventy thousand balanced samples whose labels record the presence of floating material.

3. Scope and Assumptions

The first assumption comes from the task definition. The paper studies two-dimensional structural topology optimisation with the design domain discretised on a 64 by 64 pixel grid. Conclusions at this resolution do not transfer directly to three dimensions or higher resolution: the paper itself lists extension to 3D and to higher resolution as future work, and notes that because topology optimisation updates every pixel of the design domain, computational complexity grows sharply with resolution. The paper also does not use a named standard benchmark such as an MBB beam; the task is constructed by randomly combining volume fractions, displacement boundary conditions and loads, so its conclusions are not conclusions about a classical benchmark.

The second assumption comes from how the data was built. The training set uses forty-two displacement boundary conditions, and testing introduces five further conditions unseen in training, for forty-seven in total. Volume fractions range from 0.3 to 0.5 in steps of 0.02, and loads are applied at randomly chosen unconstrained nodes on the design-domain boundary with direction angles from zero to pi in steps of pi over six. The main dataset holds thirty-three thousand 64 by 64 samples, each with six channels: the optimal topology, the prescribed volume fraction, the von Mises stress, the strain energy density, and the load constraints in the x and y directions. It splits into thirty thousand for training, two hundred for validation, eighteen hundred for level 1 testing and one thousand for level 2 testing.

One piece of accounting deserves honest reporting: dataset samples carry six channels, while the final conditional diffusion model is described as taking four channels, namely volume fraction, strain energy density, von Mises stress and boundary loads. The text does not explain further how six training channels are arranged into four model inputs. The finite element mesh size is likewise not disclosed, so the 64 by 64 pixel grid should not be equated with a finite element resolution.

The third assumption comes from the evaluation design. The level 1 test set uses the forty-two seen boundary conditions with unseen constraint combinations, while the level 2 test set uses five boundary conditions absent from training, also with unseen combinations. This layering is the prerequisite for reading every number in the paper, because the two factor claims appear on different levels and point at different metrics within the same level.

The fourth assumption concerns what the metrics physically mean. Compliance error is defined as the relative difference between the compliance of the generated topology and that of the SIMP ground truth; a positive value means the generated structure is less stiff than the reference, and a negative value means it exceeds it. The paper gives two explicit reasons for not using pixel error as the final metric: the SIMP ground truth need not be strictly optimal because of penalisation and filtering, and even a tiny pixel error can accompany a large compliance error when key locations lack material. Floating material is a boolean that is one when the topology contains floating material; load violation is a boolean that is one when no material exists where a load is applied. The paper also reports volume fraction error and the proportion of samples whose compliance error exceeds thirty percent.

The fifth assumption is an easily overlooked runtime fact. The paper reports 0.06 seconds per generated topology for TopologyGAN and 21.59 seconds for TopoDiff, a difference of about 360 times. It lists accelerated sampling as a direct route to practicality, which also means that when evaluating this technology, the premise that generative methods are faster does not hold for the original implementation and has to be measured against a real workflow.

4. Getting Started

The path from zero to reproducing TopoDiff results has three stages, each with checkable acceptance points.

In preparation, confirm three things: the Python environment and deep learning framework versions, the code repository, which the authors publish on GitHub as francoismaze/topodiff, and how the data is obtained. The main dataset, the two surrogate datasets and the test set are distributed through a Dropbox link given on the project page, and the trained models sit behind the same link. These are third-party hosted endpoints whose availability has to be verified by actually accessing them, and should not be assumed permanently reachable.

In the data stage, first get the data generation chain working. If the goal is only to reproduce the published results, use the released datasets. Extending them, whether by widening the volume-fraction range, adding boundary-condition types or raising resolution, requires reproducing the SIMP and finite element pipeline, since the authors' own dataset was generated with a Python SIMP implementation. Reproducibility here depends on three things: the penalisation factor and filtering settings, the finite element solver details, and the random sampling seed. Missing any one of them makes regenerated data incomparable with the released data.

In the model stage, proceed with the surrogates before the diffusion model, because guidance depends on their gradients. Train the compliance regressor and the floating-material classifier first. Their validation accuracy directly determines guidance quality: if a surrogate predicts poorly on noisy topologies, guidance pushes denoising in the wrong direction. The paper tunes hyperparameters on a validation set of two hundred structures by grid search, and the core knobs are the regression guidance scale, the classification guidance scale, and the maximum noise level at which each guidance term is enabled. Each candidate configuration is evaluated by actually generating a topology and running a finite element evaluation, because the effect of guidance strength cannot be inferred from offline surrogate accuracy.

In the inference and evaluation stage, follow the paper's protocol. Each test condition is run nine times before averaging, and that step cannot be skipped because single-sample variance is large enough to hide differences between methods. Evaluate level 1 and level 2 separately and report them separately. After generation, run a finite element evaluation to obtain the true compliance rather than reusing the surrogate prediction, since the surrogate prediction is precisely what guidance acted upon and reusing it amounts to self-verification. For statistical significance, use paired tests against TopologyGAN on exactly the same test set.

At the point of engineering integration, decide in advance what role the method plays. At 21.59 seconds per sample it is unsuitable as the real-time engine of an interactive design tool. It fits two uses: generating initial guesses for traditional iterative optimisation, or acting as a filter over a large batch of candidates, where a generative model proposes many geometries and finite element checks discard the infeasible ones. These two uses pull accuracy and speed in opposite directions, and the choice belongs in the proposal rather than being discovered later.

5. Comparison with Alternatives

The spectrum of topology-optimisation methods can be arranged along three axes: whether the method iterates, whether it learns, and whether the objective explicitly contains physical quantities.

Traditional SIMP-type methods iterate, do not learn, and optimise compliance and constraints directly, at the cost of a full sensitivity iteration for every design. Their risk of suboptimality comes from penalisation and post-processing, which is a different mechanism entirely from the failure modes of generative methods, since the latter have no explicit optimisation process and only imitate the distribution of ground truths.

TopologyGAN, by Nie and colleagues in the Journal of Mechanical Design in 2021, is the baseline used in this paper. It feeds multiple physical fields computed on the original unoptimised material domain into the generator of a conditional GAN, reducing the mean squared error by nearly three times and the mean absolute error by 2.5 times against a baseline cGAN on test problems with previously unseen boundary conditions, and introduces a hybrid U-SE-ResNet generator. Choosing it as the baseline is defensible: it also uses physical-field conditioning, also performs generative topology optimisation, and also published its implementation and trained networks.

A conditional variational autoencoder is the third baseline adopted in the 3D extension paper, with a loss combining mean squared reconstruction error and a KL regularisation term. The conclusions about TopoDiff rest on comparisons at matched conditions against these baselines.

Follow-up work belongs in the same judgement. A 2025 preprint on latent space diffusion for topology optimisation recasts the pixel-space diffusion of TopoDiff as a variational-autoencoder latent diffusion: a VAE compresses the image space to reduce the parameters and computation needed for sampling, the latent space is sampled from a standard normal so that nearby topologies blend continuously, and floating material, load imbalance and volume fraction deviation are moved into auxiliary training losses so that the two auxiliary predictors operating on noisy images can be dropped. That paper reports improvements over existing diffusion-based methods in compliance accuracy, volume control and structural connectivity.

On the 3D side, a 2025 article in the Journal of Applied Physics pushes the conditional diffusion and surrogate guidance scheme to three-dimensional 3D printing, with a dataset of eighty-five thousand structures split into seventy thousand training, five thousand validation and ten thousand test, covering forty-five boundary conditions and fifty predefined boundary scenarios, with volume fractions sampled uniformly between 0.2 and 0.8. The network input is a 33 by 17 by 9 tensor with ten channels and the output design domain is 32 by 16 by 8 with values between zero and one, and each auxiliary model is trained on thirty thousand samples. The reported mean compliance error is 3.23 percent for diffusion against 36.28 percent for a WGAN-GP baseline and 41.11 percent for a VAE baseline, with floating material proportions of 5.03, 41.94 and 52.11 percent respectively.

In the broader physics-AI landscape, topology optimisation belongs to the generate-with-constraint-guidance family, alongside diffusion routes for materials, inverse problems and geometry generation. Its intersections with this site are the surrogate modelling collection, the multiphysics coupling collection, the industrial pipeline entry on OpenFOAM, which describes where such generative models connect to a high-fidelity solver stack, and the cost boundary guide, whose ledgers determine whether a generate-then-filter combination pays off.

6. Benchmarks and Evidence

The quantitative conclusions have to be read level by level, because reading them together yields the opposite of what the paper says.

On level 1, the in-distribution boundary conditions with unseen constraint combinations, mean compliance error is 48.51 percent for TopologyGAN, 4.10 percent for TopoDiff without guidance and 4.39 percent with guidance; median compliance error is 2.06, 0.80 and 0.83 percent; the proportion with compliance error above thirty percent is 10.11, 2.33 and 2.56 percent; mean volume fraction error is 11.87, 1.86 and 1.85 percent; load violation is zero for all three; and floating material proportions are 46.78, 6.64 and 5.54 percent.

On level 2, the out-of-distribution boundary conditions, mean compliance error is 143.08 percent for TopologyGAN, 22.13 percent without guidance and 18.40 percent with guidance; median compliance error is 6.82, 1.88 and 1.82 percent; the proportion above thirty percent is 24.10, 8.20 and 8.10 percent; mean volume fraction error is 14.31, 1.81 and 1.80 percent; load violation is again zero for all three; and floating material proportions are 67.90, 7.53 and 6.21 percent.

The two factors in the abstract and conclusion follow from level 2. Physical performance error goes from 143.08 to 18.40 percent, a ratio of 7.78, roughly eight; floating material goes from 67.90 to 6.21 percent, a ratio of 10.93, roughly eleven. On level 1 the directions invert: compliance error goes from 48.51 to 4.39 percent, a ratio of 11.05, and floating material from 46.78 to 5.54 percent, a ratio of 8.44. The text on level 1 says that average compliance error improves by about eleven times and floating material by more than eight, which is exactly swapped relative to the abstract.

The ablation on the guidance strategy itself gives a more restrained conclusion. On level 1, mean compliance error with and without guidance is 4.39 against 4.10 percent with a paired two-tailed test giving p = 0.1, that is, no significant change; floating material is 5.54 against 6.64 percent. On level 2, mean compliance error is 18.40 against 22.13 percent, a relative reduction of about seventeen percent with a paired one-tailed p = 0.05, and floating material 6.21 against 7.53 percent, about eighteen percent. Against TopologyGAN on level 2, the paired one-tailed test for the mean compliance error reduction gives p = 9 times ten to the minus twelve, and for mean volume fraction error p = 5 times ten to the minus one hundred and sixty. The paper's reading is that guidance mainly reduces the number of samples with extremely high compliance error rather than significantly shifting the median.

Taken together, these numbers support a specific judgement: the gain from replacing the GAN with a diffusion model greatly exceeds the gain from guidance, and the value of guidance concentrates on out-of-distribution cases and on manufacturability, with no significant effect on mean compliance error in the in-distribution regime. Describing TopoDiff as a method that uses surrogate guidance to simultaneously improve mechanical performance goes beyond what the paper's own evidence supports.

The 3D extension carries the same kind of accounting caveat: its abstract claims roughly eight times lower average physical performance error and eleven times fewer infeasible samples, while the direct ratios in the body of the paper are, against the WGAN-GP baseline, mean compliance error of 36.28 against 3.23 percent (about 11.2 times) and floating material of 41.94 against 5.03 percent (about 8.3 times), and against the VAE baseline, 41.11 against 3.23 percent (about 12.7 times) and 52.11 against 5.03 percent (about 10.4 times). The text also reports that classifier guidance reduces floating material occurrence by eighty-two percent. One passage uses 8,500 test samples while the dataset split states ten thousand test samples against a total of eighty-five thousand, which is not fully self-consistent.

7. Role in Physics-AI Workflows

Within physics AI, TopoDiff exemplifies the generate-with-external-guidance paradigm, sharing a lineage with diffusion models applied to materials, inverse problems and geometry generation. Three transferable lessons stand out.

The first is to replace sparse inputs with physical fields. Encoding loads and boundary conditions as dense von Mises stress, strain energy density and volume fraction instead of sparse vectors improved the learnability of the conditioning. The lesson is not limited to topology optimisation: any generative task with sparse inputs and dense outputs can benefit by rewriting the input as a physically sufficient field that the simulator produces directly.

The second is to treat the surrogate as a source of guidance rather than as an evaluator. The two surrogates are frozen after training and supply gradients during sampling, injecting the optimisation objective into denoising. This is a general interface for combining a generative model with an optimiser: any differentiable physical estimator can serve as the guidance source. The later latent-space work moves floating material and load mismatch into training-time auxiliary losses, representing another solution to the same problem, internalising guidance at the cost of more complex training and less flexibility.

The third is to define metrics before choosing methods. Separating physical performance, manufacturability and constraint satisfaction into compliance error, floating material, load violation and volume fraction error stops visual plausibility and structural usability from being conflated. That separation transfers directly to other generative engineering-design tasks.

Within a data-production chain, the role of TopoDiff must be stated plainly: it is a candidate generator and coarse filter, not a final deliverable. A sound workflow has the generative model propose many candidate geometries, finite element and manufacturability checks filter them, and survivors enter detailed design. One candidate needs one finite element evaluation, which is tractable on a 64 by 64 grid, so a flow that generates a thousand candidates, evaluates a thousand and keeps twenty is realistic. What is genuinely constrained is interactive use, where a 21.59 second per-sample cost rules it out.

The relationship with other material on this site is complementary rather than duplicative. The cost boundary guide supplies the ledgers for judging when learned models actually pay off, and topology optimisation is one concrete case for those ledgers: the denominator here is not a full transient solve but one topology generation plus one finite element evaluation. The surrogate modelling collection places generative methods within a wider spectrum, the industrial OpenFOAM pipeline entry covers the role of a high-fidelity solver as reference and fallback, and the building energy surrogate case is another instance of the same generate-then-filter structure.

8. Limitations and Failure Modes

The first class of failure concerns resolution and dimensionality. The core experiments run on a 64 by 64 two-dimensional grid, and the 3D extension drops the output to a 32 by 16 by 8 voxel resolution. Real industrial topology optimisation typically operates in three dimensions at far higher resolution. Treating conclusions from 64 by 64 as evidence of three-dimensional feasibility is a scale error, and the 3D extension paper itself lists resolution as an ongoing challenge.

The second class concerns surrogate reliability. Guidance injects gradients from two surrogates into denoising, so a surrogate that predicts poorly on noisy topologies steers generation the wrong way. Such failure does not appear as a crash but as results that look reasonable and perform slightly worse, which is hard to detect inside a pipeline. The sound practice is to recalibrate surrogates periodically against high-fidelity evaluation and to treat guidance strength as a scenario-dependent hyperparameter rather than a fixed setting.

The third class concerns the data boundary. Training covers forty-two boundary conditions, and testing with five out-of-distribution conditions already sits at the edge of what the model generalises to. The boundary-condition space of real engineering problems is far larger, spanning different fixing arrangements, contact constraints and multi-case load combinations, all outside the training distribution. The 18.40 percent compliance error on level 2 shows that out-of-distribution error is much better than the baseline yet still far from engineering-ready in absolute terms.

The fourth class is a mismatch of cost and process. The 360-fold inference gap rules TopoDiff out of interactive design tools and of real-time optimisation needing many candidates. It fits instead as an initial-guess generator for traditional SIMP: let the diffusion model supply a high-performance, structurally connected starting point and let gradient iteration refine it locally. That division lets each component do what it is good at, with the generative model giving global structure and the gradient method giving local optimality.

The fifth class is misuse of the evaluation. The most dangerous is substituting the surrogate prediction for a finite element evaluation, which is self-verification; the other is mixing factor claims across test levels, since the same paper gives factors pointing in opposite directions on different levels. Any unqualified statement of an N-fold performance gain is a misreading of exactly these numbers. A defensible report names the level, the metric definition, the baseline and the sample count.

A further class of failure lies outside the model, in downstream use. A generated layout matching the reference in compliance does not match it in fatigue life, vibration modes, creep or manufacturing process constraints. Compliance is a single static quantity, whereas real structures are determined by several physical quantities at once. When the deliverable must be certified, a topology-optimisation result still needs a full engineering verification path.

9. Version and Source Notes

Record 1: the identity of the paper. The earlier entry pointed its source URL at the arXiv page and recorded the licence as an arXiv preprint with code and data released. The formal identity is Mazé and Ahmed's Diffusion Models Beat GANs on Topology Optimization, published in the AAAI 2023 proceedings, volume 37, pages 9108 to 9116, with arXiv v1 submitted on 20 August 2022 and v2 revised on 5 December 2022 under a CC BY 4.0 arXiv licence. The year field should read 2023 for the formal publication, with 2022 being the first public preprint.

Record 2: author affiliation. The arXiv page lists a single author affiliation, the Massachusetts Institute of Technology, for François Mazé and Faez Ahmed, consistent with their MIT email addresses and the project page credit. Mazé uses a Mines Paris PSL address and the paper acknowledges the Mines Paris Foundation and MISTI France, but the text does not list Mines Paris as a formal affiliation. The earlier phrasing "from MIT" holds at the institutional level, with the more precise description being the MIT DeCoDe Lab.

Record 3: the scope of the factor claims. The abstract and conclusion figures of roughly eight times lower average physical performance error and roughly eleven times fewer infeasible samples refer to the level 2 out-of-distribution test set, with compliance error from 143.08 to 18.40 percent and floating material from 67.90 to 6.21 percent. On level 1 the ratios invert, with compliance error from 48.51 to 4.39 percent and floating material from 46.78 to 5.54 percent. The earlier entry omitted the level, which is a missing piece of accounting rather than a wrong number.

Record 4: the magnitude of the guidance contribution. The ablation shows no significant difference between guided and unguided mean compliance error on level 1, with 4.39 against 4.10 percent and p = 0.1. On level 2 guidance reduces mean compliance error by about seventeen percent with p = 0.05 and floating material by about eighteen percent. The bulk of the method's benefit comes from substituting the diffusion model for the GAN, while the incremental contribution of guidance concentrates in out-of-distribution settings and on manufacturability. The earlier description attributed the method's behaviour to surrogate guidance without separating these two contributions.

Record 5: inference cost. The paper reports 0.06 seconds per generated topology for TopologyGAN against 21.59 seconds for TopoDiff, about 360 times, and identifies accelerated sampling as a direct route to practicality. The earlier entry omitted this, leaving the efficiency narrative of generative methods replacing iterative optimisation without its key denominator.

Record 6: task configuration. The paper studies two-dimensional topology optimisation on a 64 by 64 pixel grid, with the task built from randomly combined volume fractions between 0.3 and 0.5 in steps of 0.02, forty-two training boundary conditions plus five out-of-distribution ones, and load direction angles from zero to pi in steps of pi over six, with no named standard benchmark such as an MBB beam. The main dataset has thirty-three thousand six-channel samples, the final diffusion model takes four channels, and the finite element mesh resolution is not disclosed. The earlier entry omitted all of these settings.

Record 7: metric definitions. Compliance error is the relative difference between generated and reference compliance; floating material is a boolean that is one when the topology contains floating material; load violation is one when no material exists where a load is applied; volume fraction error is the relative deviation of the generated volume fraction. Load violation is zero for every model on both test sets, so the reduction in infeasible samples in the abstract comes in practice from floating material. The earlier wording of infeasible samples corresponds in the results tables to the floating material proportion.

Record 8: baseline identity. The comparison baseline is TopologyGAN by Nie and colleagues, and the 3D extension paper adds a conditional VAE. In the 3D paper's table the direct ratios, against the WGAN-GP baseline at 36.28 against 3.23 percent compliance error and 41.94 against 5.03 percent floating material, do not fully match the abstract's eightfold and elevenfold wording, so any transcription of the 3D conclusions needs the same qualification.

Record 9: disambiguating a same-named method. A 2025 Nature Machine Intelligence article from a Tsinghua group uses the name TopoDiff for a global-geometry-aware latent encoding framework for protein backbone generation, which has no technical connection to the structural topology optimisation method discussed here and shares only the name. Searching and citing needs the distinction.

Record 10: data and code availability. The datasets and trained models are distributed through a Dropbox link on the project page and the code lives in the francoismaze/topodiff repository on GitHub. The long-term availability of third-party hosted links has to be verified by actual access and should not be assumed stable in a proposal.

All records were checked on 11 November 2026 against the arXiv:2208.09591 v2 full text and abstract, the AAAI publication record, the author project page, the Journal of Applied Physics article, the abstract of the latent-space preprint, and the journal record for TopologyGAN. Every quoted figure names its test level, metric definition and baseline, and no figure is merged across levels.

10. References

  1. Mazé F., Ahmed F., "Diffusion Models Beat GANs on Topology Optimization", arXiv:2208.09591v2 (revised 5 December 2022, CC BY 4.0; checked 11 November 2026).
  2. Mazé F., Ahmed F., "Diffusion Models Beat GANs on Topology Optimization", Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37, pp. 9108-9116, 2023, DOI 10.1609/aaai.v37i8.26093.
  3. MIT DeCoDe Lab, "TopoDiff project page: datasets, code, trained models and evaluation protocol", decode.mit.edu/projects/topodiff (checked 11 November 2026).
  4. Nie Z., Lin T., Jiang H., Kara L. B., "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.
  5. Bekbolat A., Kurokawa S., Zaman F. H. K., Kamal M. A. S., Shehab E., Ali M. H., "Diffusion models for topology optimization in 3D printing applications", Journal of Applied Physics, 137(13), 135102, 2025, DOI 10.1063/5.0246189.
  6. Lutheran A., Das S., Tabarraei A., "Latent Space Diffusion for Topology Optimization", arXiv:2508.05624 (submitted 7 August 2025).
  7. Rombach R. et al., "High-Resolution Image Synthesis with Latent Diffusion Models", CVPR 2022, arXiv:2112.10752 (latent diffusion foundation).
  8. Ho J., Jain A., Abbeel P., "Denoising Diffusion Probabilistic Models", NeurIPS 2020, arXiv:2006.11239 (variance schedule and the fixed-covariance result).
  9. Sigmund O., "A 99-line Python code for compliant cantilever beam topology optimization", GitHub (SIMP reference implementation).
  10. Sigmund O., Maute K., "Topology optimisation: A finite element-based homogenisation approach", Structural and Multidisciplinary Optimization, 20(3), 2003.
  11. DClaw Research, "The Cost Boundary of Classical Solvers: When Physics-AI Surrogates Actually Pay Off in 2026", site guide (11 November 2026).
  12. DClaw Research, "Surrogate modelling and reduced-order modelling collection", site collection (11 November 2026).
Tags: