acceptodds
Under review as a conference paper at ICLR 2027

TopoMamba: A Load-Support Relation-Guided Multi-Directional State-Space Model for Topology Optimization

Abstract

Deep learning has emerged as an efficient alternative for predicting high-performance material distributions in topology optimization. Existing methods struggle to accurately capture load-transfer information, limiting out-of-distribution generalization, while their model architectures often incur high computational costs. To address these challenges, this paper proposes TopoMamba, a topology prediction framework incorporating a load-support relation-guided multi-directional state-space model. Coupling physical fields with load-support relations enables more effective modeling of mechanical dependencies. A load-support relation-guided spatially adaptive fusion mechanism dynamically adjusts multi-directional scan features according to spatial conditions. Mamba is coupled with the solid isotropic material with penalty method to enhance structural mechanical performance while maintaining computational efficiency. Results on two-dimensional topology optimization benchmarks demonstrate that TopoMamba achieves superior topology prediction accuracy, out-of-distribution generalization, and computational efficiency over state-of-the-art models. The proposed load-support physics-guided framework enables efficient optimization of more complex structural systems. Codes are available at https://anonymous.4open.science/r/anonymous-ICLR-TM.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Topology Optimization, Mamba, Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_4Vv7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TopoMamba predicts two-dimensional material-density fields using finite-element physical fields, explicit load–support features, and a four-direction Mamba backbone. Its distinctive mechanism is a spatially varying scalar gate between horizontal and vertical scan aggregates; predictions subsequently undergo optional SIMP refinement. The model is evaluated on the TopoDiff/NITO 2D benchmark: 30k training samples, 1,800 in-distribution (ID) and 1,000 out-of-distribution (OOD) test samples. The paper reports compliance error (CE), volume-fraction error (VFE), parameter count and end-to-end runtime, plus a single-seed ablation. The headline claims are 0.17% mean CE on ID and 1.07% on OOD with 10 SIMP steps, using 7M parameters and 0.48 s per design.

Critical strengths

TopoMamba has a reasonable motivation that use the relationship between loads and supports to guide how the model combines information from different spatial directions. The compact model and the reported results after SIMP refinement are encouraging. However, the current experiments leave substantial uncertainty about both the reliability of the predictions and the benefit of the proposed architectural changes that I have tried to outline in the weaknesses section below.

Critical weaknesses

It seems to me that the method distinction is the specific load–support representation and its use in a spatial scalar blend between row and column scan groups. However, Table 4 reports favorable but small absolute changes from the proposed relation and fusion components, without uncertainty estimates or sufficiently precise ablation definitions. These results neither establish negligible effects nor conclusively establish robust improvements. The paper consequently offers a promising engineering result without yet explaining how much its distinctive mechanisms contribute.

The ablations provide limited evidence for the distinctive contribution. Removing the load–support relations or adaptive fusion changes the reported errors much less than removing the inherited physical-field inputs, and no variation across training runs is reported. These smaller gains may be useful, but they need a controlled comparison with a standard four-direction backbone using the same inputs and training procedure, evaluated before and after refinement. The method description also needs to explain how relations defined at support nodes become features at every patch, since that operation is central to reproducing the proposed mechanism.

Physical-field conditioning is explicitly inherited from TopologyGAN, as secgtion 3.3 already acknowledges in your paper. Four-route visual state-space scanning is established by the cited VMamba. Learned prediction followed by a few SIMP steps is already central to NITO. So contribution (3), described as a prediction-optimization collaborative framework, is not independently new as presently specified and the claims should be clearly stated.

TO-MambaLDM already combines Mamba, physical conditioning, and topology generation, although for a different multiphysics cooling-design problem. Optimize Any Topology (OAT) addresses the same benchmark family and reports better mean CE than this paper’s baseline NITO. It is important to discuss both works and qualify the broad state-of-the-art claim by resource regime. A comparison with the OAT architecture is needed to establish performance claims, as it offers additional resolution and shape-dependency benefits.

The evaluation excludes predictions with compliance errors above 1000% without reporting how many predictions are removed for each method. Table 2 omits unrefined TopoMamba and OOD volume-fraction error, preventing assessment of the predictor’s standalone OOD performance and constraint satisfaction. Because refinement precedes binarization, the prescribed volume constraint during refinement alone does not establish feasibility of the final binary output; the thresholding procedure and OOD constraint results matter.

Overall, the reporting makes it difficult to judge robustness, since a low average among retained predictions can coexist with frequent failures. The paper should report failure rates, constraint violations, and results on common test conditions. It should also explain exactly what changes in the out-of-distribution test set and report accuracy and runtime at matching refinement budgets.

Several comparisons also mix refinement budgets (Tables 1–2). Such comparisons can describe complete systems, but cannot isolate the predictor’s advantage.

Questions for authors

See questions mentioned above.

8 comments
AuthorsOpenReview ↗

Response to Weaknesses 1: Contribution of Relation Encoding and Adaptive Fusion

R1. Contribution of Relation Encoding and Adaptive Fusion

We thank the reviewer for identifying the specific methodological distinction: the load–support representation and its use in conditional blending of row- and column-scan features. To address concerns regarding effect sizes and ablation rigor, we provide a rigorous four-way controlled evaluation holding physical fields, the Cross4 backbone, and training settings strictly constant across three independent runs (reporting mean and sample standard deviation in Tables R3-1 and R3-2).

While the absolute performance gains over the already strong baseline are nuanced, the ablation results demonstrate systematic improvements in key regimes. Specifically, the full model consistently enhances initial prediction quality and median performance: relative to removing both modules, it reduces ID mean CE by up to 10.66% at 0 steps and decreases OOD median CE by 8.46% and 6.91% at 0 and 5 steps, respectively. This confirms that explicit relation encoding and adaptive fusion provide a superior starting point for downstream physical refinement.

As physical SIMP iterations increase (5 and 10 steps), iterative optimization naturally compensates for initial differences, causing performance gaps to narrow. Minor variations in specific metrics (such as OOD 0-step mean CE or 10-step median CE) transparently illustrate the boundaries of these modules—primarily serving as an effective initialization framework rather than replacing physical update

Table R3-1. Four-way controlled ablation with matched physical inputs and Cross4 backbone.

DatasetConfiguration0 steps mean CE (%)0 steps median CE (%)5 steps mean CE (%)5 steps median CE (%)10 steps mean CE (%)10 steps median CE (%)
IDFull model2.97 ± 0.150.29 ± 0.010.43 ± 0.070.13 ± 0.000.17 ± 0.020.07 ± 0.00
IDWithout relation-encoding branch3.13 ± 0.290.29 ± 0.010.50 ± 0.050.14 ± 0.000.20 ± 0.040.08 ± 0.00
IDFixed equal-weight fusion3.34 ± 0.120.34 ± 0.040.47 ± 0.010.14 ± 0.000.21 ± 0.040.08 ± 0.00
IDBoth modules removed3.32 ± 0.280.32 ± 0.020.48 ± 0.020.14 ± 0.010.19 ± 0.010.08 ± 0.00
OODFull model15.62 ± 0.641.79 ± 0.072.02 ± 0.100.32 ± 0.031.20 ± 0.050.17 ± 0.01
OODWithout relation-encoding branch16.33 ± 0.901.92 ± 0.152.00 ± 0.170.35 ± 0.031.20 ± 0.180.18 ± 0.02
OODFixed equal-weight fusion16.40 ± 1.621.90 ± 0.162.05 ± 0.120.34 ± 0.001.20 ± 0.040.17 ± 0.01
OODBoth modules removed15.41 ± 1.411.96 ± 0.162.07 ± 0.090.35 ± 0.011.23 ± 0.060.16 ± 0.00

Table R3-2. Joint-module effect across SIMP budgets.

DatasetSIMPmean CE: Both modules removed → Full modelRelative changemedian CE: Both modules removed → Full modelRelative change
ID03.32 ± 0.28 → 2.97 ± 0.15↓10.66%0.32 ± 0.02 → 0.29 ± 0.01↓9.72%
ID50.48 ± 0.02 → 0.43 ± 0.07↓8.85%0.14 ± 0.01 → 0.13 ± 0.00↓3.46%
ID100.19 ± 0.01 → 0.17 ± 0.02↓7.38%0.08 ± 0.00 → 0.07 ± 0.00↓1.25%
OOD015.41 ± 1.41 → 15.62 ± 0.64↑1.32%1.96 ± 0.16 → 1.79 ± 0.07↓8.46%
OOD52.07 ± 0.09 → 2.02 ± 0.10↓2.26%0.35 ± 0.01 → 0.32 ± 0.03↓6.91%
OOD101.23 ± 0.06 → 1.20 ± 0.05↓2.00%0.16 ± 0.00 → 0.17 ± 0.01↑4.43%
AuthorsOpenReview ↗

Response to Weaknesses 2: Controlled Ablations and Node-to-patch Mapping

R2. Controlled Ablations and Node-to-patch Mapping

We thank the reviewer for identifying the specific methodological distinction. To directly address concerns regarding effect, uncertainty, and mechanism clarity, we combine controlled ablation studies with mechanism .

First, to quantify each mechanism, we evaluate four configurations while holding the physical-field inputs, the scanning backbone, and all training settings strictly constant across three independent runs (Tables R3-1 and R3-2). Notably, the two-module ablation uses a standard four-direction backbone. While the absolute gains over the strong physical-field baseline are moderate, the results reveal systematic improvements in initial prediction quality and median compliance: relative to removing both modules, the full model reduces ID mean CE by 10.66%, 8.85%, and 7.38% at 0, 5, and 10 SIMP steps, and lowers OOD median CE by 8.46% and 6.91% at 0 and 5 steps. As SIMP iterations progress, physical refinement naturally narrows the gap, transparently showing that the primary value of these modules lies in providing a superior structural initialization.

Second, these quantitative gains are directly corroborated by our mechanism representation analyses. Specifically, linear probes on frozen features demonstrate that the relation-encoding branch successfully decodes explicit geometric relations, achieving an R2R^2 of 0.67±0.010.67 \pm 0.01 in predicting FEA-derived critical load-bearing regions across the OOD set (substantially outperforming raw inputs at R2=0.50R^2 = 0.50). Furthermore, our analysis of the adaptive fusion module confirms that the network dynamically tailors directional blending to specific loading and boundary conditions predominantly via layer-wise weight modulation, rather than collapsing to uniform averaging. Together, the controlled ablations and mechanistic analyses demonstrate that both modules perform their intended physical roles.

Finally, we will clarify the exact node-to-patch mapping and feature integration pipeline in the revised manuscript:

  1. Node-Grid Representation: On the 65×6565 \times 65 finite-element node grid, we construct two seven-channel groups corresponding to the xx- and yy-displacement constraints. Each group encodes a load-presence indicator, two relative-displacement components, two normalized force components, a signed geometric moment, and squared distance, yielding 14 channels masked by the active supports.

  2. Element-Level Aggregation: The relation CNN processes these inputs into 64-channel node features, and each 64×6464 \times 64 element feature is computed by averaging its four corner-node features.

  3. Feature Integration & Patch Embedding: A learned scalar gate scales the relation feature before adding it to the base conditioning feature, after which a 4×44 \times 4 patch embedding generates 256 spatial tokens for the Mamba backbone.

  4. Geometry-Conditioned Routing: A separate routing path generates layer- and patch-specific row/column blending weights. Thus, relation encoding shapes the initial spatial representations, whereas adaptive fusion governs internal directional mixing during sequence propagation

AuthorsOpenReview ↗

Response to Weaknesses 3: Relation to Prior Components and Contribution

R3. Relation to Prior Components and Contribution

We thank the reviewer for pointing out the inaccuracies in our contribution statements. We do not claim that the combination of neural prediction and SIMP optimization is proposed here for the first time. Instead, we wish to clarify the original intent of Contribution 3, which aims to highlight the practical value of TopoMamba within a complete topology design pipeline: delivering high-quality initial material distribution via fast prediction, followed by a small number of SIMP iterations to further enhance mechanical performance, thereby achieving a favorable trade-off between structural quality and computational efficiency.

Accordingly, we have revised Contribution 3 in the manuscript as follows:

TopoMamba enables high-quality structural designs with only a few SIMP refinement steps, achieving a favorable balance between structural performance and computational efficiency.

Reviewer_rxqrOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

TopoMamba predicts two-dimensional material-density layouts from physical fields and load–support features using a four-direction Mamba backbone, followed by optional SIMP refinement. The paper reports improved compliance error and runtime on ID and OOD benchmarks.

Critical strengths

None.

Critical weaknesses

W1. No convincing substantive contribution beyond image-based regression.

The method remains supervised reconstruction of SIMP-generated density images on two-dimensional regular grids. Introducing Mamba and relation-conditioned fusion does not, by itself, establish a new optimisation principle or a meaningful extension of solvable problems. Woldseth et al. [1] critically question direct-design prediction and distinguish image reconstruction from solving the physical optimisation problem. Although this paper adds SIMP refinement and is therefore a hybrid method, it does not convincingly resolve that underlying concern. Even if its reported speedup is valid, speed alone does not establish sufficient methodological novelty or scientific significance for acceptance.

W2. The experiments do not demonstrate a decisive advance.

The learning loss does not directly optimize compliance. Physical-field conditioning does not enforce mechanical equilibrium in the predicted structure, while SIMP subsequently optimises its physical performance. In Table 4, removing relations or adaptive fusion increases OOD mean CE only from 1.79% to 1.83% or 1.91%, without reported uncertainty. Moreover, cases with CE above 1000% are excluded without per-method failure counts. These results do not establish a robust, independently useful contribution from the proposed architecture.

Questions for authors

Q1. What substantive capability does this method add beyond reproducing existing SIMP solutions and supplying a learned initialisation? How does it address the direct-design limitations discussed by Woldseth et al. [1]?

Q2. Do the claimed benefits persist with failure counts included, repeated-run uncertainty, and comparisons against classical SIMP and simple warm starts at matched solution quality, including data-generation and training costs?

[1] Woldseth, R. V., Aage, N., Bærentzen, J. A., and Sigmund, O. (2022). On the use of artificial neural networks in topology optimisation. Structural and Multidisciplinary Optimization, 65, 294. https://doi.org/10.1007/s00158-022-03347-1

1 comment
AuthorsOpenReview ↗

Tables 2-1 in Response R1

Table R2-1. Controlled structural-performance ablation (common valid ID/OOD cases). All configurations retain identical physical-field inputs and the Cross4 backbone; three trained runs are evaluated for each.

DatasetConfiguration0 steps mean CE (%)0 steps median CE (%)5 steps mean CE (%)5 steps median CE (%)
IDFull model2.97 ± 0.150.29 ± 0.010.43 ± 0.070.13 ± 0.00
IDWithout relation-encoding branch3.13 ± 0.290.29 ± 0.010.50 ± 0.050.14 ± 0.00
IDFixed equal-weight fusion3.34 ± 0.120.34 ± 0.040.47 ± 0.010.14 ± 0.00
IDBoth modules removed3.32 ± 0.280.32 ± 0.020.48 ± 0.020.14 ± 0.01
OODFull model15.62 ± 0.641.79 ± 0.072.02 ± 0.100.32 ± 0.03
OODWithout relation-encoding branch16.33 ± 0.901.92 ± 0.152.00 ± 0.170.35 ± 0.03
OODFixed equal-weight fusion16.40 ± 1.621.90 ± 0.162.05 ± 0.120.34 ± 0.00
OODBoth modules removed15.41 ± 1.411.96 ± 0.162.07 ± 0.090.35 ± 0.01
Reviewer_6wFDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors propose TopoMamba, which tries to incoporate load-support relation guidance in multi-directional state space models along with better fusion architecture. Finally, they also provide a better prediction-optimization collaborative framework for topology optimization. They demonstrate superior performance in terms of accuracy, out-of-distribution generalization, and computational efficiency.

Critical strengths
  • the idea of prediction-optimization collaborative framework is really nice.
  • Similarly, the idea of using mamba based architecture along with fusion is very good.
  • Significant improvement in the performance, specifically in the case of OOD generalization is nice.
  • Results show much better performance than other baselines considered.
Critical weaknesses
  • A lot of ideas are presented, but not thoroughly validated in the results. Some ablations are there, but none that completely satisfies the claims made.
  • The details of what is considered OOD is not clear to me. What did you do differently to generate the OOD samples. Without these details how are we supposed to know.
  • The claim in the abstract that this framework being enabler for efficient optimization for more complex systems is not proven anywhere. In fact, in the conclusion, it is stated particularly that the experimentation on complex systems is left for future work. With such a limitation, this claim should be reconsidered.
  • Similarly, the authors claim collaborative framework. For a framework to be collaborative, there should be transfer of information from both sides the optimization and the prediction. It is unclear to me how the optimization is affecting the prediction here.
  • The degree of OOD-ness and the degree of collaboration etc. are not well addressed for the authors to sufficiently claim these as contributions of the paper.
Questions for authors

Please see above. At minimum, I would like to know how are the authors describing OOD-ness and would like to tone-down their claims. This would help make the paper much better in terms of presentation.

Comments on the paper

1 comment

The author-reviewer discussion period closes on November 18 (AoE). Please engage during this period so that we have our issues discussed thoroughly.

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.