acceptodds
Under review as a conference paper at ICLR 2027

Learning Topology-Aware Representations for Mechanics-Relevant TPMS Metamaterials

Abstract

Learning informative representations of mechanical Triply Periodic Minimal Surface (TPMS) metamaterials is important for mechanics analysis and structural design, yet existing methods have rarely explicitly supervised learned representations with information about global structural connectivity that is relevant to mechanical response. We propose a topology-aware representation learning framework that integrates persistent homology into a DMAE backbone to explicitly encode global structural connectivity. A persistence-guided masking strategy prioritizes patches associated with persistent topological features, while an auxiliary persistent-homology decoder encourages the latent representation to retain information about topological feature counts across connected components, tunnels, and voids. We evaluate the learned representations on two complementary TPMS datasets using family classification, mechanical-property classification, unsupervised clustering, and regression. Across the two datasets, our method demonstrates strong performance in both structural family and mechanical-property classification, improving family classification by up to 3.9% while achieving the best linear-probe performance. More notably, unsupervised property-clustering performance improves by up to 26.7% and 51.5% across the two datasets, while strong regression performance is retained with . These results establish persistent topology as an effective inductive bias for representation learning of complex three-dimensional mechanical metamaterials.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
3D representation learning, AI4Science, Mechanical metamaterials, Persistent homology, Triply periodic minimal surfaces (TPMS)

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_VXCuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a topology-informed self-supervised model for predicting the mechanical properties of triply periodic minimal-surface metamaterials. A 3D denoising masked autoencoder with a ViT backbone is pretrained on 64364^3 signed-distance fields, after which its frozen representations are evaluated on mechanical-property tasks. Persistent homology modifies the model in two ways: patch-masking probabilities are weighted by the persistence of local H0H_0, H1H_1, and H2H_2 features, and an auxiliary head predicts the log feature count in each homology dimension from the pooled latent, with weight λPD\lambda_{\mathrm{PD}}. The motivation is that small topological changes can produce much larger changes in effective stiffness, as illustrated by the cut/merge example in Figure 1. Experiments use 13,114 MetaDB structures and 18,549 interpolated TPMS structures, with homogenized stiffness tensors as labels. The learned representations are tested on family classification, C11C_{11} classification, unsupervised C11C_{11} clustering, and C11C_{11} regression against nine baselines: MVAE, SVAE, VICReg, I-JEPA, MAE, DMAE, ColorMAE, DaM, and ScalableTopo. The clearest gains are in unsupervised clustering, where η2\eta^2 rises from 0.27 to 0.41 on Interpolation and from 0.15 to 0.32 on MetaDB. Family classification improves on MetaDB, while regression remains comparable to DMAE. The appendix includes component ablations, sensitivity sweeps, and a latent-alignment analysis.

Critical strengths
  1. The motivation is clear, and the empirical results are strong. The teaser figure cut/merge example shows that a change of less than 1% in volume can alter the persistence diagram while changing stiffness by 7–10%. The experiments cover two datasets, nine baselines, four downstream tasks, a five-seed ablation with standard deviations, and a hyperparameter sweep. The paper also reports cases where other methods perform better, including the Interpolation linear probes and MetaDB regression.

  2. The PH objective produces a clear clustering gain. For unsupervised C11C_{11} clustering, η2\eta^2 increases from 0.272 to 0.412 on Interpolation and from 0.147 to 0.323 on MetaDB. These improvements are larger than the variation across seeds and are reflected in both AccMatch and macro-F1 in Table S3. Appendix A.4 also describes the persistent-homology pipeline in enough detail to make the experiment reproducible.

Critical weaknesses
  1. It is difficult to tell which part of the method is new, and the ablation does not support the paper’s explanation. The paper has no related-work section, so Table 1 is the only comparison. According to the table, PH-based supervision is already used by TopoAE (Moor et al., 2020), Hofer et al. (2019), and PH-SSL (Luo et al., 2023), however these methods are not discussed in the text. This seems to leave persistence-guided masking as the main new component. Table S5 does not show a measurable benefit from that component. On Interpolation, the PH head alone reaches η2=0.405±0.03\eta^2=0.405\pm0.03, compared with 0.412±0.040.412\pm0.04 for the full model. Masking alone reaches 0.251, below the 0.272 baseline. On MetaDB, the three results, 0.315, 0.315, and 0.323±0.0130.323\pm0.013, are effectively indistinguishable. The PH head also reduces regression performance there, from R2=0.937R^2=0.937 without it to 0.899 with it. I therefore do not see evidence for the claimed “complementary roles” of the two components. The method notation also needs to be cleaned up. The total loss appear as λPHLCount\lambda_{\mathrm{PH}}L_{\mathrm{Count}}, λPDLPH\lambda_{\mathrm{PD}}L_{\mathrm{PH}}, and λPDLtopo\lambda_{\mathrm{PD}}L_{\mathrm{topo}}, with an additional λeff\lambda_{\mathrm{eff}} introduced in Appendix A.10. The masking ratio is denoted by both ρ\rho and rr. No values are given for the dimension weights ωq\omega_q, and the persistence threshold ε=0.005\varepsilon=0.005, which determines the regression targets, appears only in Appendix A.4.
  2. The evaluation protocol is not clear enough to support the reported results. Appendix A.10 says that hyperparameters were selected using downstream-task metrics, but does not say whether those metrics came from the validation split. This matters because the results are quite sensitive to the chosen values: η2\eta^2 ranges from 0.097 to 0.461 across λPD\lambda_{\mathrm{PD}} and from 0.150 to 0.317 across ρ\rho. The paper also does not explain how the nine baselines were tuned. There is also an unresolved discrepancy between tables. The selected configuration obtains η2=0.461/0.613/0.610\eta^2=0.461/0.613/0.610 in Table S6, but the corresponding value in Tables S3 and S5 are 0.412/0.582/0.5760.412/0.582/0.576. These appear to be results for the same configuration and should either agree or be clearly distinguished. Since the PH head predicts three feature counts, Tasks 2–4 should include a probe using only [log⁡n0,log⁡n1,log⁡n2,volume fraction].[\log n_0,\log n_1,\log n_2,\text{volume fraction}]. This would show the source of the downstream gains. Finally, several figures make the comparisons harder to assess. Figure 5 connects four unrelated descriptors with lines, Figure 3 places ten methods on one radar chart, and Figure 4 groups different metrics from the two datasets under a shared label. Separate plots with explicit axes would make the results easier to verify.
Questions for authors

None

10 comments
AuthorsOpenReview ↗

Response 1/7: Novelty, related work, and introduction of additional baselines

Reviewer #3 (VXCu) — Response 1/7

We sincerely thank the reviewer for the time and effort spent on our manuscript and for the constructive feedback, which has improved it substantially. The comments on novelty led us to add a related-work discussion and a TopoAE baseline, which clarify how our approach differs from existing PH-based methods. The suggested probe on PH counts and volume fraction, together with the closer reading of the ablation, helped us describe the roles of the two mechanisms more accurately: their contributions are dataset-dependent, and the full model is the most consistent across both datasets. We have also clarified the evaluation protocol, unified the notation, and redrawn Figures 3–5 as suggested. We address each point below.

Reviewer comment:

It is difficult to tell which part of the method is new, and the ablation does not support the paper’s explanation. The paper has no related-work section, so Table 1 is the only comparison. According to the table, PH-based supervision is already used by TopoAE (Moor et al., 2020), Hofer et al. (2019), and PH-SSL (Luo et al., 2023), however, these methods are not discussed in the text. This seems to leave persistence-guided masking as the main new component.

Response: We thank the reviewer for pointing out these relevant works. In the original manuscript, we included the related-work discussion within the Introduction. To make the comparison with prior studies clearer, we have added a separate related-work section discussing these methods in the revised manuscript. We have also expanded it to include TopoAE, Hofer et al., and PH-SSL, and to clarify the differences in motivation and methodology.

For motivation, TopoAE (Moor et al., 2020) aims to preserve topological relationships between the input data and their latent representations; Hofer et al. (2019) focuses on controlling latent-space connectivity for one-class learning; PH-SSL introduces persistent topological fingerprints to capture more fine-grained relationships between molecules.Our work is motivated by the relationship between structural connectivity and mechanical response in 3D TPMS metamaterials. Local changes in connectivity can alter load-transfer pathways and effective stiffness, even when the overall geometric variation is small. As illustrated by the controlled Cut/Merge experiments in Fig. 1 and Appendix Sec. A.1, localized structural modifications produce noticeable changes in both persistent-topological descriptors and directional mechanical responses.

And for methodology, these approaches incorporate PH into representation learning in different ways. TopoAE uses a topological loss to preserve relationships between the input and latent spaces; Hofer et al. (2019) introduces PH-based regularization to control latent-space connectivity; PH-SSL uses persistent topological fingerprints to provide self-supervision through topology prediction and contrastive learning. Our method instead incorporates PH into two stages: First, persistence-guided masking assigns higher masking probabilities to patches associated with persistent topological features, encouraging the encoder to learn their structural context. Second, an auxiliary PH prediction head supervises the latent representation using the counts of persistent features in H0H_0, H1H_1, and H2H_2. Together, the two mechanisms encourage learning representations to capture structural connectivity relevant to mechanical response.

We’ve updated the explanations for the topology informed methods in Related Work:

(lines 091-099) “Prior topology-aware methods, as summarized in Table 1, have largely incorporated PH to encode structural connectivity in 2D (Hu et al., 2023; Wang et al., 2020a) or in generic geometric settings, through topological pooling and representations (Chen & Gel, 2023; Tola et al., 2025). Several works further use PH directly as a training signal: TopoAE (Moor et al., 2020) and Hofer et al. (2019) preserve the topology of data or latent manifolds, and PH-SSL (Luo et al., 2023) uses persistence-based fingerprints for molecular self-supervision. These methods, however, are not designed around mechanical response.”

To demonstrate that these mechanisms are well suited to this objective, we added TopoAE and PH-SSL as baseline comparisons to test whether their topology preservation methods are sufficient, and we also considered a direct PH-descriptor (log⁡(H0)\log(H_0), log⁡(H1)\log(H_1), log⁡(H2)\log(H_2), volume fraction) baseline to test whether simple geometric features alone are sufficient. Overall, our method achieves stronger performance than the three topology-informed baselines across most mechanics-related metrics on both datasets.

(Continue in Response 2/7)

AuthorsOpenReview ↗

Response 2/7: Baseline analysis, four result tables, and manuscript revisions

Reviewer #3 (VXCu) — Response 2/7

(Continue from Response 1/7)

These results support two main observations. First, our proposed mechanisms are more effective for mechanics-oriented representation learning in most evaluated metrics. For example, on MetaDB, TopoAE performs well in family classification (kNN macro-F1 = 0.978) but achieves a C11C_{11} clustering η2\eta^2 of only 0.012, compared with 0.323 for our model. This may suggest that preserving topological relationships alone does not necessarily produce representations that capture mechanical-property variations.

Second, directly using PH descriptors does not provide consistently strong performance across different tasks. For example, on MetaDB, PH Descriptor achieves a C11C_{11} regression R2R^2 of 0.423, compared with 0.899 for our model. Although PH Descriptor achieves a higher clustering η2\eta^2 on Interpolation TPMS (0.492 versus 0.412), our method performs better on the corresponding clustering accuracy and macro-F1, as well as the classification and regression tasks. These results, together with our ablation studies, suggest that integrating PH information into the learning process provides broader benefits than directly using PH descriptors alone.

Specifically, persistence-guided masking directs the encoder toward regions associated with persistent topological features, while the auxiliary PH head encourages the learned representation to retain global topological information. Together, these mechanisms incorporate PH into the learning of full 3D geometry representations, with the goal of capturing structural characteristics relevant to mechanical response.

The result tables (Task1-Task4) for comparison between the three baselines and our proposed Full model are shown below:

Task1 - Family Classification (new results were updated to Table 2 in the revised manuscrip)

ModelMetaDB F Lin ↑MetaDB F kNN ↑Interpolation Linear Strict Acc ↑Interpolation Linear Relaxed Acc ↑Interpolation kNN Strict Acc ↑Interpolation kNN Relaxed Acc ↑
TopoAE0.9610.9780.7500.7740.7720.897
PH Descriptor0.2900.5210.0130.0230.2100.261
PH-SSL0.9880.9740.8720.8810.8300.939
Ours0.9940.9850.8670.8800.8310.939

Task 2 -C11C_{11} Classification (new results were updated to Table S2 in the revised manuscript)

ModelMetaDB kNN Bal. Acc ↑MetaDB kNN Macro-F1 ↑MetaDB Linear Bal. Acc ↑MetaDB Linear Macro-F1 ↑Interpolation kNN Bal. Acc ↑Interpolation kNN Macro-F1 ↑Interpolation Linear Bal. Acc ↑Interpolation Linear Macro-F1 ↑
ScalableTopo0.9540.9540.9600.9570.8560.8560.9160.916
TopoAE0.9500.9500.9300.9300.8250.8250.8770.877
PH Descriptor0.6660.6650.4430.4140.8390.8400.8430.844
PH-SSL0.9360.9360.9420.9420.8710.8710.9120.912
Ours0.9510.9510.9640.9640.8870.8870.9240.925

Task 3 -C11C_{11} Clustering (new results were updated to Table S3 in the revised manuscript)

ModelMetaDB η2\eta^2 ↑MetaDB AccMatch ↑MetaDB Macro-F1 ↑Interpolation η2\eta^2 ↑Interpolation AccMatch ↑Interpolation Macro-F1 ↑
ScalableTopo0.1990.5430.5330.1580.4220.395
TopoAE0.0120.4710.4640.1880.4760.477
PH Descriptor0.0260.3810.3060.4920.5540.513
PH-SSL0.0230.5930.5860.2860.5150.502
Ours0.3230.6600.6730.4120.5820.576

Task 4 - C11C_{11} Regression (new results were updated to Table S4 in the revised manuscript)

ModelMetaDB MAE ↓MetaDB RMSE ↓MetaDB R2R^2 ↑Interpolation MAE ↓Interpolation RMSE ↓Interpolation R2R^2 ↑
ScalableTopo0.0080.0410.9320.0030.0050.992
TopoAE0.0420.0920.6850.0040.0060.987
PH Descriptor0.0460.1240.4230.0100.0140.934
PH-SSL0.0100.0470.9170.0030.0050.992
Ours0.0090.0500.8990.0030.0050.992

Corrections:

We have added the following paragraph to Section 3.2 (Baselines) of the revised manuscript:

(lines 328-337) “Additionally, we include three topology-informed baselines: ScalableTopo, which uses scalable persistent-homology regularization; TopoAE, which preserves topological information in learned representations; and PH-SSL, which incorporates persistent-homology-based self-supervision to learn topology-aware representations. We also include PH-based descriptors (H0H_0, H1H_1, and H2H_2 persistence counts, together with volume fraction), which directly use persistent-homology feature counts as structural descriptors.”

The corresponding baseline results have also been added to Table 2 and Tables S2–S4 in the revised manuscript.

AuthorsOpenReview ↗

Response 3/7: Roles of the two Mechanisms and Ablation Interpretation

Reviewer #3 (VXCu) — Response 3/7

Reviewer comment:

Table S5 does not show a measurable benefit from that component. On Interpolation, the PH head alone reaches η2=0.405±0.03\eta^2 = 0.405 \pm 0.03, compared with 0.412±0.040.412 \pm 0.04 for the full model. Masking alone reaches 0.251, below the 0.272 baseline. On MetaDB, the three results, 0.315, 0.315, and 0.323±0.0130.323 \pm 0.013, are effectively indistinguishable. The PH head also reduces regression performance there, from R2=0.937R^2 = 0.937 without it to 0.899 with it. I therefore do not see evidence for the claimed “complementary roles” of the two components.

Response: Thank you for highlighting the benefit for each of our proposed component. We agree that, on a given metric, the full model is often only marginally better than the stronger single component, and that the PH head reduces C11C_{11} regression on MetaDB (0.937 → 0.899). We have therefore revised the term “complementary roles” to describe more precisely what Table S5 shows: the two mechanisms contribute in a dataset-dependent way, and their combination removes the need to choose one per dataset.

On Interpolation TPMS, the gain comes mainly from the PH head. Adding the head to the baseline raises clustering η2\eta^2 from 0.272 to 0.405, whereas masking alone slightly lowers it (0.251). On MetaDB, masking provides benefits that the head does not. Masking alone raises η2\eta^2 from 0.147 to 0.315, gives the best family kNN macro-F1 (0.993), and gives the best C11C_{11} regression R2R^2 (0.937), compared with 0.907 for the head-only model. Thus, the most effective single component differs between the two datasets: head-only ranks second on Interpolation TPMS (average rank 2.06) but third on MetaDB (2.63), while masking-only ranks last on Interpolation TPMS (3.38) but second on MetaDB (2.00).

The full model does not dominate every entry, but it is the most consistent across both datasets: it achieves the best mean on 12 of the 16 metric–dataset entries and the best average rank on both datasets (1.38). We have revised the text to present this as a dataset-dependent tradeoff rather than as uniform complementarity.

Corrections:

(Lines 427-428) In Section 3.4: we revised the conclusion of ablation study to “their advantages are dataset-dependent”

(Lines 477-483) In Section 4: Ablation studies show that the contributions of the two topology-aware components are dataset-dependent: the PH head contributes more on Interpolation TPMS, whereas persistence-guided masking contributes more on MetaDB. Combining both components removes the need to select one per dataset, and the full model achieves the most consistent performance across both datasets.

(Lines 533-539) “Nevertheless, our study has two main limitations that need further investigations. First, the persistent-homology targets are computed on the finite voxel domain of each sample without periodic identification, so topological features that close only across the periodic boundary are not counted; a periodic filtration would align more directly with the periodic mechanical connectivity of TPMS structures. Second, physically meaningful descriptors such as volume fraction strongly determine axial stiffness, and how to combine such descriptors with learned topology-aware representations remains an open direction for future work.”

Reviewer comment:

The method notation also needs to be cleaned up. The total loss appears as λPHLCount\lambda_{\mathrm{PH}}L_{\mathrm{Count}}, λPDLPH\lambda_{\mathrm{PD}}L_{\mathrm{PH}}, and λPDLtopo\lambda_{\mathrm{PD}}L_{\mathrm{topo}}, with an additional λeff\lambda_{\mathrm{eff}} introduced in Appendix A.10.

Response: We apologize for the confusion caused by the inconsistent notation. We have standardized the notation for the total training objective throughout the manuscript as:

Ltotal=Lrec+λPDLPHL_{\mathrm{total}}=L_{\mathrm{rec}}+\lambda_{\mathrm{PD}}L_{\mathrm{PH}}

where LrecL_{\mathrm{rec}} denotes the reconstruction loss, LPHL_{\mathrm{PH}} represents the persistent-homology supervision loss, and λPD\lambda_{\mathrm{PD}} controls its contribution.

Corrections:

(Lines 179-181) Figure 2 (Page 4, Line 180): Changed "Ltotal=Lrec+λPHLCountL_{\mathrm{total}}=L_{\mathrm{rec}}+\lambda_{\mathrm{PH}}L_{\mathrm{Count}}" to "Ltotal=Lrec+λPDLPHL_{\mathrm{total}}=L_{\mathrm{rec}}+\lambda_{\mathrm{PD}}L_{\mathrm{PH}}".

(Lines 179-181) Appendix A.3 (Page 19, Line 995): Changed "Ltotal=Lrec+λPDLtopoL_{\mathrm{total}}=L_{\mathrm{rec}}+\lambda_{\mathrm{PD}}L_{\mathrm{topo}}" to "Ltotal=Lrec+λPDLPHL_{\mathrm{total}}=L_{\mathrm{rec}}+\lambda_{\mathrm{PD}}L_{\mathrm{PH}}".

(Lines 179-181) Appendix A.10 (Page 28, Line 1477): "the PH-loss weight λPD\lambda_{\mathrm{PD}}, which are most relevant ".

These corrections ensure consistent notation across the manuscript without changing the actual training objective or experimental implementation. We thank the reviewer for pointing out these inconsistencies.

Reviewer_86ysOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposed a topology-aware representation learning framework that integrates persistent homology into a DMAE backbone to explicitly encode global structural connectivity. A persistence-guided masking strategy prioritizes patches associated with persistent topological features, while an auxiliary persistent-homology decoder encourages the latent representation to retain information about topological feature counts across connected components, tunnels, and voids.

Critical strengths

The paper correctly identifies that existing representation learning methods for metamaterials rarely explicitly supervise learned representations with global topological information, which is fundamentally relevant to mechanical response. The physical motivation, that global structural connectivity (tunnels, cavities, void channels) governs effective elastic properties, is clearly articulated and well-supported by the perturbation analysis in Appendix.

Critical weaknesses

The computational cost is not analyzed. The paper compares with ScalableTopo but does not discuss or compare with other recent topology-aware representation learning approaches. This limits the positioning of the work within the broader TDA literature. In Table 2, while the proposed method's performance on the Interpolation TPMS dataset reveals notable weaknesses. The paper acknowledges this in the text ("the linear-probe results remain slightly below the best baseline") but does not provide a detailed analysis of why the topology-aware representation fails to fully capture family-relevant structural information under linear probing on the interpolation dataset. Given that the method's core claim is about preserving topology-relevant information, this gap between kNN and linear probe performance on Interpolation TPMS warrants further discussion.

Questions for authors

Provide a more detailed analysis of why the proposed method underperforms relative to baselines on Interpolation TPMS linear-probe metrics (Table 2) and on MetaDB regression metrics (Table S4), and discuss whether the topological supervision contributes meaningfully beyond what is achieved by advanced masking strategies alone.

6 comments
AuthorsOpenReview ↗

Response 1/4: Computational Cost and Topology-Aware Baseline Comparisons

Reviewer #2 (86ys) — Response 1/4

Reviewer comment 1:

The computational cost is not analyzed.

Response: We evaluate the computational cost of the proposed topology-aware mechanisms using the same NVIDIA H100 GPU configuration (40GB MIG instance). Compared with vanilla DMAE, the full model increases the parameter count by only 4.27% (from 3.5085M to 3.6584M), while the typical training time increases by approximately 2.25% (from 69.68 to 71.25 seconds per epoch). Training time increases only slightly because persistent homology is computed offline, and the resulting topological information is cached to avoid repeated computations during training. During downstream inference, all variants have identical inference FLOPs (21.754G). Overall, the proposed method adds little training overhead and no additional inference computation.

ModelParams (M) ↓Training (s/epoch) ↓Inference (ms/sample) ↓Inference FLOPs (G) ↓
Vanilla DMAE3.508569.682.361± 0.07421.754
Count Head Only (w/o Mask)3.658470.682.380 ± 0.12021.754
PD Mask Only (w/o Head)3.508571.082.865 ± 0.03321.754
Full Model (Ours)3.658471.252.881 ± 0.03121.754

Corrections:

(Lines 1674-1700) We have also added a new subsection, “Computational Cost Analysis” (Appendix A.12), together with a computational cost comparison table.

Reviewer comment 2:

The paper compares with ScalableTopo but does not discuss or compare with other recent topology-aware representation learning approaches.

Response: We thank the reviewer for this suggestion. We added three new baselines: TopoAE and PH-SSL to test whether their topology preservation methods are sufficient, and we also considered a direct PH-descriptor (log⁡(H0)\log(H_0), log⁡(H1)\log(H_1), log⁡(H2)\log(H_2), volume fraction) baseline to test whether simple geometric features alone are sufficient. Overall, our method achieves stronger performance than the three topology-informed baselines across most mechanics-related metrics on both datasets.

These results support two main observations. First, our proposed mechanisms are more effective for mechanics-oriented representation learning in most evaluated metrics. For example, on MetaDB, TopoAE performs well in family classification (kNN macro-F1 = 0.978) but achieves a C11C_{11} clustering η2\eta^2 of only 0.012, compared with 0.323 for our model. This may suggest that preserving topological relationships alone does not necessarily produce representations that capture mechanical-property variations.

Second, directly using PH descriptors does not provide consistently strong performance across different tasks. For example, on MetaDB, PH Descriptor achieves a C11C_{11} regression R2R^2 of 0.423, compared with 0.899 for our model. Although PH Descriptor achieves a higher clustering η2\eta^2 on Interpolation TPMS (0.492 versus 0.412), our method performs better on the corresponding clustering accuracy and macro-F1, as well as the classification and regression tasks. These results, together with our ablation studies, suggest that integrating PH information into the learning process provides broader benefits than directly using PH descriptors alone.

Specifically, persistence-guided masking directs the encoder toward regions associated with persistent topological features, while the auxiliary PH head encourages the learned representation to retain global topological information. Together, these mechanisms incorporate PH into the learning of full 3D geometry representations, with the goal of capturing structural characteristics relevant to mechanical response.

(Continue in Response 2/4)

AuthorsOpenReview ↗

Response 2/4: Additional Baseline Results Across Four Downstream Tasks

Reviewer #2 (86ys) — Response 2/4

(Continue from Response 1/4)

The result tables (Task1-Task4) for comparison between the three baselines and our proposed Full model are shown below:

Task1 - Family Classification (new results were updated to Table 2 in the revised manuscrip)

ModelMetaDB F Lin ↑MetaDB F kNN ↑Interpolation Linear Strict Acc ↑Interpolation Linear Relaxed Acc ↑Interpolation kNN Strict Acc ↑Interpolation kNN Relaxed Acc ↑
TopoAE0.9610.9780.7500.7740.7720.897
PH Descriptor0.2900.5210.0130.0230.2100.261
PH-SSL0.9880.9740.8720.8810.8300.939
Ours0.9940.9850.8670.8800.8310.939

Task 2 -C11C_{11} Classification (new results were updated to Table S2 in the revised manuscript)

ModelMetaDB kNN Bal. Acc ↑MetaDB kNN Macro-F1 ↑MetaDB Linear Bal. Acc ↑MetaDB Linear Macro-F1 ↑Interpolation kNN Bal. Acc ↑Interpolation kNN Macro-F1 ↑Interpolation Linear Bal. Acc ↑Interpolation Linear Macro-F1 ↑
ScalableTopo0.9540.9540.9600.9570.8560.8560.9160.916
TopoAE0.9500.9500.9300.9300.8250.8250.8770.877
PH Descriptor0.6660.6650.4430.4140.8390.8400.8430.844
PH-SSL0.9360.9360.9420.9420.8710.8710.9120.912
Ours0.9510.9510.9640.9640.8870.8870.9240.925

Task 3 -C11C_{11} Clustering (new results were updated to Table S3 in the revised manuscript)

ModelMetaDB η2\eta^2 ↑MetaDB AccMatch ↑MetaDB Macro-F1 ↑Interpolation η2\eta^2 ↑Interpolation AccMatch ↑Interpolation Macro-F1 ↑
ScalableTopo0.1990.5430.5330.1580.4220.395
TopoAE0.0120.4710.4640.1880.4760.477
PH Descriptor0.0260.3810.3060.4920.5540.513
PH-SSL0.0230.5930.5860.2860.5150.502
Ours0.3230.6600.6730.4120.5820.576

Task 4 - C11C_{11} Regression (new results were updated to Table S4 in the revised manuscript)

ModelMetaDB MAE ↓MetaDB RMSE ↓MetaDB R2R^2 ↑Interpolation MAE ↓Interpolation RMSE ↓Interpolation R2R^2 ↑
ScalableTopo0.0080.0410.9320.0030.0050.992
TopoAE0.0420.0920.6850.0040.0060.987
PH Descriptor0.0460.1240.4230.0100.0140.934
PH-SSL0.0100.0470.9170.0030.0050.992
Ours0.0090.0500.8990.0030.0050.992

Corrections:

We have added the following paragraph to Section 3.2 (Baselines) of the revised manuscript:

(lines 328-337) “Additionally, we include three topology-informed baselines: ScalableTopo, which uses scalable persistent-homology regularization; TopoAE, which preserves topological information in learned representations; and PH-SSL, which incorporates persistent-homology-based self-supervision to learn topology-aware representations. We also include PH-based descriptors (H0H_0, H1H_1, and H2H_2 persistence counts, together with volume fraction), which directly use persistent-homology feature counts as structural descriptors.”

The corresponding baseline results have also been added to Table 2 and Tables S2–S4 in the revised manuscript.

AuthorsOpenReview ↗

Response 3/4: Analysis of the Linear-Probe Performance Gap on Interpolation TPMS

Reviewer #2 (86ys) — Response 3/4

Reviewer comment 3:

In Table 2, while the proposed method's performance on the Interpolation TPMS dataset reveals notable weaknesses. The paper acknowledges this in the text ("the linear-probe results remain slightly below the best baseline") but does not provide a detailed analysis of why the topology-aware representation fails to fully capture family-relevant structural information under linear probing on the interpolation dataset. Given that the method's core claim is about preserving topology-relevant information, this gap between kNN and linear probe performance on Interpolation TPMS warrants further discussion.

Response: We thank the reviewer and agree that this gap deserved more than one sentence. In short, the lower linear-probe accuracy on Interpolation TPMS is a small and controlled trade-off rather than a loss of family information.Table S6 shows that this gap is governed by the PH-loss weight: it arises only on this dataset because its family labels are compositional. To support this point, we conducted a dominance-margin error analysis comparing our model with ScalableTopo, the strongest baseline on the linear probe. The analysis shows that the gap is concentrated in blends whose two largest components have nearly equal weights. In these cases, the “wrong” prediction is usually the other major component of the blend, not an unrelated family. We chose the setting(full model) because that it favours mechanics-relevant organization, which is the objective of the paper, and now state this cost explicitly in Sec. 3.3.1.

The gap is a controlled trade-off rather than a loss of family information. Relative to its own DMAE backbone, our model improves the linear probe on Interpolation TPMS from 0.724/0.852 to 0.867/0.880 (strict/relaxed); the remaining gap to the best baseline is 0.018/0.017. Table S6 shows that this gap is governed by the PH-loss weight: linear strict accuracy decreases monotonically from 0.887 to 0.837 as λPD\lambda_{\mathrm{PD}} increases from 0.01 to 0.20, while kNN accuracy is nearly unchanged. At λPD\lambda_{\mathrm{PD}} = 0.01 the linear probe is level with the best baseline (0.887/0.899 vs 0.885/0.897), but C11C_{11} clustering degrades sharply (η2\eta^2 = 0.097 vs 0.461 at the adopted λPD\lambda_{\mathrm{PD}} = 0.05). We chose the setting that favours mechanics-relevant organization, which is the objective of the paper, and now state this cost explicitly in Sec. 3.3.1.

Why the cost appears on this dataset. On Interpolation TPMS the family label is the component with the largest blending weight in a continuous design space. Structures on either side of a label boundary (e.g., 40/30/30 vs 30/40/30 blends) are nearly identical in topology and stiffness, so the label is compositional rather than topological. PH supervision organizes the latent space by topological feature counts, which vary across these boundaries continuously; local neighbourhoods remain consistent (kNN is on par with the best baseline), but the classes are less separable by a single hyperplane. On MetaDB, where families are discrete base types, the same model gives the best linear and kNN results (0.994/0.985).

To understand more on this, and the performance differences between Full and ScalableTopo, we conducted a matched-protocol error analysis across three representation seeds. We grouped test samples by dominance margin—the difference between the two largest interpolation weights—and examined whether incorrectly predicted families were present in the original blend. ScalableTopo achieved slightly higher linear strict accuracy (86.80% versus 86.15%), whereas Full achieved higher kNN strict accuracy (83.35% versus 81.27%). Samples with dominance margins ≤0.10, including ties, accounted for 96.3% of ScalableTopo’s net linear advantage and 99.4% of Full’s net kNN advantage. Furthermore, 99.22% of Full’s linear-probe strict-error records were nonempty predictions containing only families already present in the blend. These findings suggest that the observed classification differences primarily concern dominant-family assignment within mixed structures, particularly when component weights are close or tied.

(Continue in Response 4/4)

Reviewer_QW8dOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies representation learning for triply periodic minimal surface (TPMS) mechanical metamaterials. It extends a 3D denoising masked autoencoder with two mechanisms derived from persistent homology (PH): masking probabilities weighted by the persistence of features whose birth/death critical cells fall in each patch, and an auxiliary head predicting log-transformed counts of retained finite features in dimensions zero through two. PH is computed from a cubical filtration of the absolute signed-distance field. At evaluation, the unmasked encoder produces a pooled, normalized representation.

The evaluation uses 13,114 MetaDB structures and 18,549 interpolated TPMS structures, with frozen-representation probes for structural family, stiffness-group classification, clustering, and continuous stiffness regression. The strongest reported gains concern MetaDB family recognition and the organization of latent clusters by the stiffness component C11. The supplement includes five-seed component ablations, full-stiffness-tensor regression, latent-distance alignment, and hyperparameter sweeps. The intended contribution is a topology-informed inductive bias for mechanics-relevant 3D representations.

Critical strengths

S1. A coherent way to inject topology into both spatial masking and global representation learning

The two mechanisms act at different levels: persistence-weighted critical-cell locations select reconstruction challenges, while the count head encourages global information to remain decodable after pooling. The topology targets are obtained from the input geometry rather than mechanical labels. This makes the approach a plausible way to incorporate domain knowledge into pretraining. The method also has a practical deployment advantage: Appendix A.3 describes downstream encoding of the full volume without requiring PH computation at inference.

S2. Promising representation results supported by several complementary evaluations

On MetaDB, Table 2 reports a linear-probe family macro-F1 of 0.994 versus 0.957 for the strongest listed baseline. Table S3 reports clustering correlation ratios of 0.323 versus 0.255 on MetaDB and 0.412 versus 0.272 on Interpolation TPMS. The corresponding relative improvements of 26.7% and 51.5% are arithmetically consistent with the table. The paper goes beyond a single downstream predictor: Table S5 includes five-seed ablations and regression of all 21 independent stiffness components, and Figure 5 examines latent-distance alignment. This breadth is useful evidence that the proposed supervision changes the representation, even though the interpretation and consistency of some results need attention.

Critical weaknesses

W1. Several evaluation results cannot be reconciled with the stated definitions or with other tables

The largest claimed advantage is in mechanical-property clustering, but its protocol is not specified sufficiently to interpret all results. Table S3 reports a MetaDB MVAE matched accuracy of 0.044. If Eq. (17) matches three clusters to the three Low/Medium/High classes on the evaluated samples, the optimal permutation accuracy must be at least 1/3: averaging over all six permutations already gives 1/3. If a different number of clusters or a different matching/evaluation procedure was used, those choices need to be specified, including how unmatched clusters are handled. The current text leaves this unresolved.

There are also inconsistencies beyond that baseline. Table S4 and the main text report Interpolation C11 R2 = 0.993 for the full method, whereas Table S5 reports a five-seed mean of 0.9923, which rounds to 0.992. Under the stated fixed test split and a common target scale, the MetaDB RMSE/R2 pairs in Table S4 also require reconciliation; in particular, DMAE and the full method have nearly identical mean R2 but different RMSE values, and the small R2 standard deviations in Table S5 do not explain this difference under ordinary per-seed averaging.

Finally, the motivating perturbation experiment has conflicting volume data. Table S1 increases occupied voxels from 162,042 for the baseline to 180,465 for Merged, approximately +11.37%, and occupancy from 18.315% to 20.398%. Figure S1 instead reports a relative volume change of -1.240% for Merged. Different geometries, resolutions, or occupancy definitions could explain this, but none is identified. These discrepancies may be reporting errors; until reconciled, they limit confidence in both the numerical comparisons and the controlled topology-mechanics argument.

W2. The added value of persistence-guided masking, and its proposed mechanical interpretation, are not sufficiently isolated

Table S5 supports a benefit from topology-related supervision, but provides weaker evidence for the necessity of combining both mechanisms. On Interpolation TPMS, the head-only model already reaches clustering eta2 = 0.4045 +/- 0.0317, compared with 0.4116 +/- 0.0363 for the full model; masking alone reaches 0.2508 +/- 0.0328, below DMAE's 0.2721 +/- 0.0060. The head-only model also has slightly higher C21 regression R2 than the full model, 0.9964 versus 0.9962. On MetaDB, adding the head to the masking-only model lowers C11 regression R2 from 0.9370 to 0.8991 and family kNN macro-F1 from 0.9930 to 0.9845, while the clustering improvement is small. These comparisons do not establish statistical significance in either direction without paired seed results, but they show a substantial task-dependent tradeoff.

The PH construction also warrants a narrower interpretation. Appendix A.4 uses |S|, so the topology signal is unchanged by exchanging S and -S, and its retained interval counts are not the Betti numbers of the load-bearing solid at a specified physical threshold. This does not invalidate an auxiliary geometric signal, especially since the encoder still receives signed geometry, but it makes the connection to mechanical connectivity something to demonstrate empirically. The paper includes ScalableTopo, yet lacks direct PH-descriptor and density/thickness controls that would distinguish the benefit of the learned representation from simpler geometric summaries. Comparisons with a surface-aware mask or shuffled spatial PH scores would more directly test whether locating persistent critical cells is responsible for the improvement. The mixed ablations and absence of these controls leave the main methodological contribution less established than the broad claim of complementary mechanisms suggests.

Questions for authors

Q1. Can the evaluation protocol and inconsistent numerical records be reconciled?

Please provide the clustering algorithm, number of clusters, fit/evaluation splits, and exact matching procedure underlying Table S3, explaining the 0.044 result relative to Eq. (17). Please also identify the runs and aggregation conventions behind Tables S4-S5, including the regression-probe model and common target scale, and provide per-seed metrics or predictions sufficient to reconcile the R2/RMSE discrepancies. For Figure S1 and Table S1, are the voxel counts, PH features, and mechanical responses computed from the same geometries and discretizations? Finally, what unit was used to split mutation/interpolation-related structures, and were all choices in Appendix A.10 made using validation data only? Corrected, consistently computed results and a leakage-free selection protocol that preserve the reported advantages would materially improve my assessment. These are clarification and validation requests about the current evidence.

Q2. What evidence isolates the benefit of spatial persistence-guided masking?

Under a common training budget and configurations selected on validation data, can the authors report paired seed differences for the full model versus the head-only and masking-only variants, and compare against a direct PH-descriptor baseline and a simple geometric control, such as density/thickness features or surface-aware masking? Please clarify how the chosen unsigned, finite-domain filtration relates to periodic mechanical connectivity, and whether the conclusion is sensitive to that choice. A consistent improvement attributable to persistence-based spatial selection, or a narrower claim supported by the existing tradeoffs, would materially affect my evaluation. This is intended to validate the proposed mechanism, rather than request a new downstream application.

11 comments
AuthorsOpenReview ↗

Response 1/10: Clarification of the Clustering Protocol and Evaluation Metrics

Response 1/10

We sincerely thank the reviewer (QW8d) for the time and care devoted to our manuscript. The detailed feedback has strengthened the work considerably. In particular, the requested direct PH-descriptor baseline and the paired-seed analysis gave us a much clearer understanding of the proposed mechanisms: their individual contributions are dataset- and task-dependent, whereas the full model performs consistently across both datasets without requiring a component to be selected for each one. The reviewer’s careful checking of our numerical records also helped us correct several reporting errors and state the evaluation protocol precisely. We address each point below.

Q1: Please provide the clustering algorithm, number of clusters, fit/evaluation splits, and exact matching procedure underlying Table S3,

Response: We thank the reviewer for this request and have added these details to Appendix A.7.2. For each frozen representation, the latent vectors are first standardized with a StandardScaler fitted on the training split only. We then apply K-means with K=3K=3 (n_init=100, random_state=42), fitted on the training latents only, to match the three C11C_{11} property groups (Low, Mid, High). The fitted model assigns clusters to the validation and test samples, and all reported metrics are computed on the test split (2,783 samples for Interpolation TPMS, with a 12,984/2,782/2,783 train/validation/test split). The property groups are defined by fixed C11C_{11} tertile thresholds based on statistics (C11≤0.0607C_{11}\leq 0.0607, 0.0607<C11≤0.10570.0607<C_{11}\leq0.1057, and C11>0.1057C_{11}>0.1057 for Interpolation TPMS). ARI and NMI compare the raw cluster assignments with these groups and are invariant to cluster permutation. For AccMatch\mathrm{Acc}{\mathrm{Match}} and FMacroF{\mathrm{Macro}}, each cluster is mapped to a property group using the training split only: the three clusters are ranked by the mean C11C_{11} of their training members and assigned to Low, Mid, and High in ascending order. This mapping is then fixed and applied to the test clusters. η2\eta^2 is computed on the test split as the fraction of C11C_{11} variance explained by the cluster assignments. The same procedure is used for MetaDB. Results are averaged over five representation seeds.

Changes in the revised manuscript:

  • A7-2, (Lines 1338 - 1344) “For each frozen representation, the latent vectors are standardized using statistics computed on the training split only. We then fit K-means with K=3K=3 clusters (100 random initializations, fixed random seed 42) on the training latents only, where KK matches the three C11C_{11} property groups. The fitted model is used to assign clusters to the held-out test samples, and all clustering metrics are computed on the test split (2,783 samples for Interpolation TPMS). The Low, Medium, and High property groups are the same as those used in Task 2. Results are averaged over five representation seeds.”
  • (Lines 1354 - 1355) “Here, the sums run over the test samples, with cluster assignments given by the K-means model fitted on the training split.”
  • (Lines 1360 - 1366) “We therefore define a mapping πtrain\pi_{\mathrm{train}} from cluster indices to property classes using the training split only: the three clusters are ranked by the mean continuous C11C_{11} of their training members and assigned to Low, Medium, and High in ascending order. This mapping is then fixed and applied to the test clusters. The matched clustering accuracy is computed on the test split as the class-balanced accuracy.”
  • (Lines 1372 - 1375) “Because the mapping is fixed from the training split rather than optimized on the evaluated samples, AccMatch\mathrm{Acc}_{\mathrm{Match}} is not bounded below by 1/3 and can fall below chance when the training-derived cluster ordering does not transfer to the test split or when the clusters are highly imbalanced.”
AuthorsOpenReview ↗

Response 2/10: Clustering Accuracy and Regression Corrections

Response 2/10

W1: The largest claimed advantage is in mechanical-property clustering, but its protocol is not specified sufficiently to interpret all results. Table S3 reports a MetaDB MVAE matched accuracy of 0.044. If Eq. (17) matches three clusters to the three Low/Medium/High classes on the evaluated samples, the optimal permutation accuracy must be at least 1/31/3: averaging over all six permutations already gives 1/31/3. If a different number of clusters or a different matching/evaluation procedure was used, those choices need to be specified, including how unmatched clusters are handled. The current text leaves this unresolved.

Q1: Please provide the clustering algorithm, number of clusters, fit/evaluation splits, and exact matching procedure underlying Table S3, explaining the 0.044 result relative to Eq. (17).

Response: Thank you for pointing out this inconsistency between the mathematical description and the implemented evaluation protocol. The reviewer is correct that Eq. (17), as written, implies an optimal permutation accuracy of at least 1/31/3 for three clusters and three reference classes evaluated on the same samples.

Our actual procedure differs from this definition: clusters are assigned to Low/Medium/High groups by ranking their mean C11C_{11} values on the training set. This mapping is then fixed and applied to the held-out test set, where we compute balanced accuracy and macro-F1. We do not optimize the mapping using test-set labels, so the 1/31/3 lower bound does not apply to this evaluation. We have corrected Eq. (17), the metric name, and the table caption to accurately describe the implemented procedure.

Changes in the revised manuscript:

  • (Lines 1366–1369) Corrected Eq. (17):
\sum_{g \in \{Low, Medium, High\}} \frac{1}{N_{test,g}} \sum_{i \in D_{test},\ y_i^{cls}=g} \mathbb{I}[\pi_{train}(z_i)=g]$$ **W1:** There are also inconsistencies beyond that baseline. Table S4 and the main text report Interpolation $C_{11}$ $R^2=0.993$ for the full method, whereas Table S5 reports a five-seed mean of 0.9923, which rounds to 0.992. **Response:** Thank you for highlighting this inconsistency. The correct result for the full method on Interpolation TPMS is the value reported in Table S5, $0.9923 \pm 0.0006$ (mean ± sample standard deviation over five seeds). This value was incorrectly rounded to 0.993 in the main text and Table S4; when reported to three decimal places, it should read 0.992 (Lines 1478–1479). We have corrected the main text and Table S4 accordingly, so that all tables now report the same result. **Changes in the revised manuscript:** - (Lines 028–029) Abstract: “More notably, unsupervised property-clustering performance improves by up to 26.7% and 51.5% across the two datasets, while strong regression performance is retained with $R^2=0.992$.” - (Lines 411–414) Section 3.3.1: “Our method achieves $R^2=0.899$ on MetaDB and $R^2=0.992$ on Interpolation TPMS, showing comparable performance on the latter while maintaining low prediction errors across both datasets.” - (Lines 1447–1450) Appendix A.7.3: “On the Interpolation TPMS dataset, it achieves the joint-lowest MAE (0.003) and RMSE (0.005), with an $R^2$ of 0.992, close to the highest baseline value of 0.993.” - (Lines 1459–1461) Table S3: “$\mathrm{Acc}{\mathrm{Match}}$ and $F{\mathrm{Macro}}$, evaluate agreement with the discrete property groups using a cluster-to-class mapping determined from the training split.”
AuthorsOpenReview ↗

Response 3/10: Reconciliation of MetaDB Regression Results

Response 3/10

W1: Under the stated fixed test split and a common target scale, the MetaDB RMSE/R2 pairs in Table S4 also require reconciliation; in particular, DMAE and the full method have nearly identical mean R2 but different RMSE values, and the small R2 standard deviations in Table S5 do not explain this difference under ordinary per-seed averaging. Please also identify the runs and aggregation conventions behind Tables S4-S5, including the regression-probe model and common target scale, and provide per-seed metrics or predictions sufficient to reconcile the R2/RMSE discrepancies.

Q1: Please also identify the runs and aggregation conventions behind Tables S4-S5, including the regression-probe model and common target scale, and provide per-seed metrics or predictions sufficient to reconcile the R2/RMSE discrepancies.

Response: Thank you for pointing out these inconsistencies. We carefully re-examined the original DMAE regression results, including the random seeds and the dataset splits used during evaluation. We confirmed that all experiments were conducted using five random seeds (42, 123, 2024, 2025, and 3407). For each seed, both DMAE and our Full method were evaluated on the corresponding held-out test split. So there’s no data leakage between training and testing. Specifically, we used an MLP regression probe with the architecture 192–256–128–64–1 to predict C11C_{11} from the frozen learned representations. The reported MAE, RMSE, and R² were computed on the original C11C_{11} scale, and the mean and sample standard deviation were calculated across the five seeds.

The original DMAE regression results (per seed) are summarized below:

5-Seed Results for DMAE model

SeedC11 RMSE ↓C11 MAE ↓C11 R² ↑
420.0496560.0094150.899064
1230.0470320.0084370.913832
20240.0524870.0093010.894217
20250.0541240.0098440.880156
34070.0468270.0079920.907912
Mean ± Std0.050025 ± 0.0032470.008998 ± 0.0007600.899036 ± 0.013013

Upon further inspection, we identified a transcription error in the full method’s regression results in Table S4: the reported RMSE of 0.047 is incorrect and should be approximately 0.050. With this correction, the RMSE of the full method is consistent with its R²: since R² and RMSE are computed on the same test split and the same target scale, a comparable R² implies a comparable RMSE, and the ranking of methods by RMSE now agrees with the ranking by R². Table S4 and the corresponding discussion in Appendix A.7.3 have been updated accordingly. The per-seed results for the full method are provided below:

5-Seed Results for Full Model

SeedC11 RMSE ↓C11 MAE ↓C11 R² ↑
420.0496280.0093120.899176
1230.0450780.0082410.920842
20240.0530180.0094770.892065
20250.0560590.0099670.871436
34070.0457350.0078560.912157
Mean ± Std0.049904 ± 0.0046990.008971 ± 0.0008860.899135 ± 0.019092

Changes in the revised manuscript:

  • (Lines 1478–1479) We’ve updated Table S4. The RMSE and R² values are now consistent under the same evaluation protocol, resolving the discrepancy identified by the reviewer.

  • (Lines 1447–1448) Appendix A.7.3: “As shown in Table S4, our method achieves competitive regression performance on MetaDB, with MAE = 0.009, RMSE = 0.050, and R² = 0.899.”

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.