acceptodds
Under review as a conference paper at ICLR 2027

Zero-Shot Compositional Generalization in Neural PDE Solvers via Conservative Material-Guided Routing

Abstract

Neural PDE solvers have shown strong generalization across new geometries and parameter regimes, but a distinct question remains less explored: if a model has seen different physical constituents separately during training, can it correctly combine that knowledge when those familiar constituents appear together for the first time? In this setting, the individual constituents are not new; what is new is their composition. We study this problem as zero-shot compositional generalization in heterogeneous PDEs. For latent-state transformer PDE solvers, providing material information as input features alone does not necessarily ensure that this information is used appropriately when forming latent physical states. To address this, we introduce Material-Interface Guided Transolver with Conservative Compositional Routing (CCR-MIG), which allows constituent information not only to be represented as input, but also to influence latent-state formation. Rather than perturbing the learned routing logits, CCR-MIG composes constituent-specific routing distributions with the original data-driven routing through a bounded convex mixture, and a conflict-aware gate reduces this intervention where constituent influences overlap. We evaluate the method on heterogeneous linear elasticity and steady thermal conduction. Across six seeded training runs, CCR-MIG reduces the mean relative error of the original Transolver from 0.08456 to 0.05818 in elasticity and from 0.03775 to 0.02323 in thermal conduction. Compared with a feature-matched baseline that receives the same constituent information but leaves latent-state formation unchanged, CCR-MIG further reduces error by 23.4% and 18.9%, respectively. Effect sizes favor CCR-MIG in all four primary comparisons, though six paired runs do not support a conclusive significance claim.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural PDE solvers, operator learning, compositional generalization, transformers, attention, heterogeneous materials, scientific machine learning, zero-shot generalization

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

This paper focuses on zero-shot compositional generalization for neural PDE solvers, where different material types are seen separately during training but appear together during testing. The authors propose CCR-MIG, which modifies Transolver by introducing material-guided latent-state routing with a bounded mixing mechanism and a conflict-aware gate. Experiments on 2D elasticity and thermal conduction show improved average accuracy compared with the original Transolver and a feature-matched baseline.

Critical strengths
  1. The paper considers an interesting and practically useful generalization problem for heterogeneous PDEs. I think directly using material information to guide latent-state routing, instead of only adding it to input features, is a reasonable idea.

  2. The proposed method is relatively simple and can be easily added to the existing Transolver framework. The results on both PDE problems are encouraging, and the ablation studies provide some evidence for the effectiveness of the proposed routing design.

Critical weaknesses
  1. I think the technical novelty is somewhat limited. The main idea is essentially to combine the original routing with material-guided routing using a bounded weighting factor. The theoretical bound itself is also quite straightforward. In addition, the benefit of the conflict-aware gate is not very clear, especially for the thermal conduction problem.

  2. The experiments are still somewhat limited in terms of problem complexity and generalization. Both benchmarks are based on relatively simple 2D linear, steady PDE problems, with mostly circular inclusions. It would be more convincing to see results on more complex geometries, material combinations, or other PDE systems. Also, although the average errors are improved, the differences are not statistically significant across the six random seeds.

Questions for authors
  1. Would simpler material-guided routing or gating designs achieve similar results? How important is the use of Gaussian-shaped routing distributions?

  2. The conflict-aware gate improves the elasticity results but has little benefit for thermal conduction. Could the authors explain this difference?

  3. Could the authors provide more quantitative results on different inclusion shapes, mesh resolutions, or more complex material combinations? This would help better understand the generalization ability of the method.

  4. Since the mixed-composition development set was used to select the routing design, could the authors further clarify how this affects the zero-shot generalization setting?

2 comments
AuthorsOpenReview ↗

Response to Reviewer bhaw

Response to Reviewer bhaw (1/2)

I think the technical novelty is somewhat limited. The main idea is essentially to combine the original routing with material-guided routing using a bounded weighting factor. The theoretical bound itself is also quite straightforward. In addition, the benefit of the conflict-aware gate is not very clear, especially for the thermal conduction problem.

We thank the Reviewer for this assessment. We agree that the bound is simple, and we present it as a design guarantee and not as a deep result. The main contribution is the finding, from a controlled comparison with the same constituent encoding, that where constituent information enters Physics-Attention matters. Conditioning the slice assignment instead of the node features lowers the error by 8.7–18.6% in six of seven benchmarks (p. 8, l. 417). The revised routing also has a precise form: each proposal is a Gaussian kernel over the learned slice embeddings, which makes it equivariant to slice relabelling (Proposition 1, p. 7, l. 343). The gate is addressed in our response to the second question below.

The experiments are still somewhat limited in terms of problem complexity and generalization. Both benchmarks are based on relatively simple 2D linear, steady PDE problems, with mostly circular inclusions. It would be more convincing to see results on more complex geometries, material combinations, or other PDE systems. Also, although the average errors are improved, the differences are not statistically significant across the six random seeds.

Thank you for this suggestion. The revision adds waisted and perforated geometries, two- and three-inclusion test compositions, and two nonlinear constituent laws. With eight seeds and the corrected holdout, the improvements over Transolver and Gaussian-bank routing are significant after Holm correction in all seven benchmarks (Table 1, p. 9, l. 432).

Would simpler material-guided routing or gating designs achieve similar results? How important is the use of Gaussian-shaped routing distributions?

We thank the Reviewer for this question, which led to the main change in the paper. The simpler submitted design, one fixed Gaussian band over the slice index per category, performs substantially worse: replacing it with constituent proposals lowers the error by 40–55% (8/8 seeds, Holm p = 0.023; Table 1 and Table 3, p. 23, l. 1225). Where the Gaussian is defined matters: over arbitrary slice indices it ignores the learned slice geometry, whereas over the learned slice embeddings it does not. For gating, removing conflict suppression does not hurt accuracy (Appendix M, p. 37, l. 1996). To answer the second question directly, we replaced the Gaussian-form proposal by a saturating tanh kernel that keeps all inputs, parameters and training settings (p. 24, l. 1264). On rectangular elasticity (mixed-400, eight seeds), this variant reaches 0.0246 compared with 0.0232 for CCR-MIG; CCR-MIG wins four of eight seeds, and the difference is not significant (p = 0.39). The variant is still 25% more accurate than Transolver (7/8 seeds). The gain therefore does not depend on the exact Gaussian form; in this experiment the Gaussian form mainly reduces the spread across seeds (SD 0.0012 versus 0.0037). What matters more is what the kernel is defined over: the learned slice embeddings rather than the slice indices.

The conflict-aware gate improves the elasticity results but has little benefit for thermal conduction. Could the authors explain this difference?

We thank the Reviewer for pointing this out. With the revised model, conflict suppression does not help in either problem. Removing it at test time, with identical weights, lowers the error by 5.8% for elasticity and by 10.3% for thermal conduction, in 8 of 8 seeds for both (p = 0.0078 each; p. 37, l. 1996, Table 16). Because the conflict term is zero during training, its strength receives no gradient, and the factor acts only as a fixed test-time rule that weakens the learned routing between soft and hard inclusions (Figure 15). The gains of CCR-MIG therefore come from the constituent proposals, and the paper now states this for both problems.

AuthorsOpenReview ↗

Response to Reviewer bhaw

Response to Reviewer bhaw (2/2)

Could the authors provide more quantitative results on different inclusion shapes, mesh resolutions, or more complex material combinations? This would help better understand the generalization ability of the method.

Thank you for this suggestion. For material combinations, the test sets contain SH, SSH and SHH plates. For mesh, the benchmarks use different meshes, including a coarser, non-conforming mesh on the perforated plate (1,585 nodes). When 30% of the input nodes are removed, the dominant slice is unchanged for 99.5% of the retained nodes and the prediction changes by 0.25% (Figure 8, p. 27, l. 1404); this is an input-sensitivity test and not a mesh-convergence study. For inclusion shapes, we tested the seed-123 models, trained on circular inclusions only, on 180 new plates with ellipse, kidney and blob inclusions (p. 37, l. 1972, Figure 13). CCR-MIG reaches 0.0261 against 0.0330 for Transolver (21.0% lower, 153 of 180 plates), with reductions of 17.9% for ellipses, 20.0% for kidney shapes and 25.1% for blobs. This test uses one seed, and multi-seed training on other shapes remains future work.

Since the mixed-composition development set was used to select the routing design, could the authors further clarify how this affects the zero-shot generalization setting?

We thank the Reviewer for raising this point. No mixed sample is used for training or for checkpoint selection, which is based only on the non-mixed validation set. The 200-sample mixed set served as the development set during method design, and no result reported in the paper is computed on it. The 400-sample holdout was generated with an independent seed only after the routing design had been fixed on that development set, and it was not used for training, checkpoint selection, hyperparameter tuning or design selection (p. 7, l. 370; Figure 3, p. 7, l. 333). For example, we retain the conflict-suppression strength λ_C = 0.55 even though removing suppression lowers the holdout error by 5.8% (p. 37, l. 1996). The model is therefore zero-shot with respect to its training data and its weights, while the development set did inform the design choices, which is why it is kept separate from the reported evaluation.

Reviewer_8qbNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies zero-shot compositional generalization in neural PDE solvers for heterogeneous media: each constituent class is seen separately during training, and the classes co-occur only at test time. It proposes CCR-MIG, a Transolver variant that adds continuous constituent-influence fields to the nodal features and uses them to build a material-guided slice distribution. That distribution is convexly mixed with the learned Physics-Attention routing under a bounded, conflict-aware gate, which guarantees an ℓ1 deviation of at most 2γmax from the original routing. CCR-MIG improves zero-shot compositional generalization over both Transolver and the feature-matched baseline on both PDEs.

Critical strengths
  1. The paper has a clean experimental design that isolates the question. The feature-matched Transolver gets exactly the same material inputs as CCR-MIG and differs only in routing.
  2. The paper has interesting findings regarding using the material information in the slice assignment. The capped version gives better and more consistent results across random seeds.
Critical weaknesses
  1. The evidence for improvement is statistically weak, and the comparisons are narrow. With six seeds, none of the four main comparisons survives Holm correction (p_adj = 0.75), and the raw p-values are 0.19–0.34. The averages are also pulled by single outlier runs: The 33.5% thermal gain over Transolver depends heavily on one Transolver run at 0.0733. In other seeds, Transolver beats CCR-MIG (0.0188 vs. 0.0272), and the medians are much closer. On elasticity, one Transolver seed reaches 0.0351, which beats every CCR-MIG run. The feature-matched average is inflated by a single run at 0.1257. On top of that, the only baselines are Transolver variants, no GeoTransolver, Transolver++, GNOT or graph-based solver. There is no “oracle” model trained on mixed plates to show how large the compositional gap is, and no in-distribution (non-mixed) results to show CCR-MIG doesn’t hurt the standard case. As it stands, I can’t tell whether the gain is real and meaningful or mostly the result of lower variance across seeds.

  2. The scope is too narrow. The benchmark is 2D, linear, with circular inclusions and exactly one soft plus one hard inclusion at test time.

  3. The conflict gate is presented as a core component, but its gain is not significant on elasticity (p = 0.56) and absent on heat conduction (d_z = −0.04).

Questions for authors

See weaknesses above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 8qbN

Response to Reviewer 8qbN (1/2)

The evidence for improvement is statistically weak, and the comparisons are narrow. With six seeds, none of the four main comparisons survives Holm correction (p_adj = 0.75), and the raw p-values are 0.19–0.34. The averages are also pulled by single outlier runs: The 33.5% thermal gain over Transolver depends heavily on one Transolver run at 0.0733. In other seeds, Transolver beats CCR-MIG (0.0188 vs. 0.0272), and the medians are much closer. On elasticity, one Transolver seed reaches 0.0351, which beats every CCR-MIG run. The feature-matched average is inflated by a single run at 0.1257.

We thank the Reviewer for this careful reading. Part of the submitted gain was an artifact of the test-set mismatch described above, and those numbers are withdrawn. With eight seeds and the corrected holdout, the gains no longer depend on single runs. On rectangular thermal conduction, CCR-MIG beats Transolver in 8/8 seeds, and the medians agree with the means (0.0091 versus 0.0174). On rectangular elasticity it wins 7/8 seeds (medians 0.0237 versus 0.0329). The one Transolver seed that beats CCR-MIG in its paired comparison is visible in Table 5. The feature-matched model no longer has an outlier run (largest error 0.0273). All per-seed values are in Tables 5–9, and Figure 10 shows the same seeds paired across models.

On top of that, the only baselines are Transolver variants, no GeoTransolver, Transolver++, GNOT or graph-based solver.

Thank you for this suggestion. We added GNOT, MeshGraphNet, GINO, FNO, the Galerkin Transformer, DeepONet, PointNet and an MLP under identical conditions (Table 2, p. 9, l. 445; Appendix H, p. 25, l. 1300). We did not evaluate Transolver++ or GeoTransolver. Transolver++ targets scaling to very large meshes and GeoTransolver targets geometry and boundary-condition shift, whereas this work studies the co-occurrence of constituents.

There is no "oracle" model trained on mixed plates to show how large the compositional gap is, and no in-distribution (non-mixed) results to show CCR-MIG doesn't hurt the standard case.

We thank the Reviewer for both suggestions, and we have carried out both. Oracle. We trained Transolver under the same protocol on 420 independently generated mixed plates (SH, SSH and SHH) and evaluated it on the same mixed-400 holdout (rectangular elasticity, eight seeds; p. 33, l. 1757, Table 15). Training on mixed plates lowers the Transolver error from 0.0329 to 0.0284, a compositional gap of 0.0044 (13.5%, 6/8 seeds). The mixed-trained model varies strongly across seeds (0.0202–0.0449), and this reduction is not statistically significant (p = 0.15). CCR-MIG, trained only on non-mixed plates, reaches 0.0232, which is 18.4% below the mixed-trained reference (6/8 seeds, p = 0.086). The reference has no constituent routing, so this comparison reflects both the training data and the architecture. We therefore do not claim that CCR-MIG closes the compositional gap. To separate the training data from the architecture, we also trained CCR-MIG on the same mixed plates. It reaches 0.0184, 20.7% below CCR-MIG trained on non-mixed plates (8/8 seeds, p = 0.0078), so a compositional gap remains for CCR-MIG as well. Under the same mixed training, however, CCR-MIG is 35.3% more accurate than Transolver (0.0184 versus 0.0284, 8/8 seeds, p = 0.0078). The benefit of routing is therefore not limited to the zero-shot setting.

In-distribution accuracy. On the non-mixed validation set of rectangular elasticity (eight seeds), the errors are 0.0250 for Transolver, 0.0170 for CCR-MIG and 0.0156 for the feature-matched model. CCR-MIG is therefore 32% more accurate than Transolver in distribution. Feature conditioning is the most accurate on compositions seen in training, but its advantage over Transolver shrinks from 38% to 23% under constituent co-occurrence, whereas that of CCR-MIG is nearly preserved (32% and 29%). This paragraph is at p. 32, l. 1718.

As it stands, I can't tell whether the gain is real and meaningful or mostly the result of lower variance across seeds.

We appreciate this concern. The improvement is in the error itself and not only in its variance. The paired per-seed reductions have confidence intervals that exclude zero against Transolver and Gaussian-bank routing in every completed benchmark, as listed in our response to Reviewer fDpY. On the two rectangular linear benchmarks, the medians also improve (0.0237 versus 0.0329 for elasticity and 0.0091 versus 0.0174 for thermal conduction).

AuthorsOpenReview ↗

Response to Reviewer 8qbN

Response to Reviewer 8qbN (2/2)

The scope is too narrow. The benchmark is 2D, linear, with circular inclusions and exactly one soft plus one hard inclusion at test time.

We thank the Reviewer for this comment. The mixed test sets now contain two- and three-inclusion plates (SH, SSH and SHH in the ratio 2:1:1). We added waisted plates with voids, a perforated plate, Fung strain-stiffening hyperelasticity (median deviation from the linear solution 9.9%) and temperature-dependent conduction. The conduction law is only weakly nonlinear, and we state this in the paper (p. 9, l. 477; p. 19, l. 1013). The benchmarks remain two-dimensional, and training uses circular inclusions; a single-seed test on non-circular inclusions is described at p. 37, l. 1972.

The conflict gate is presented as a core component, but its gain is not significant on elasticity (p = 0.56) and absent on heat conduction (d_z = −0.04).

We thank the Reviewer for raising this, and we tested it directly. No training sample contains both soft and hard constituents, so the conflict term is zero throughout training and its strength receives no gradient. Removing it at test time, with identical weights, lowers the error on the rectangular plate by 5.8% for elasticity (95% CI [3.3, 8.5]%, 8/8 seeds, p = 0.0078) and by 10.3% for thermal conduction (95% CI [5.7, 13.7]%, 8/8 seeds, p = 0.0078). We conclude that the gains come from the constituent proposals and not from conflict suppression, and the paper now says so (Appendix M, p. 37, l. 1996; Table 16; Figure 15).

Reviewer_fDpYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents CCR-MIG, improving neural PDE solvers on material combinations not seen during training. Built on Transolver, it uses continuous material influence fields to guide routing, limits the strength of this guidance, and reduces it where material influences overlap. Experiments on elasticity and steady-state heat conduction report relative error reductions of 17.4% and 33.5%, respectively, when training on single-material configurations and testing on mixed-material ones.

Critical strengths

The paper addresses a practical generalization problem. Models trained on individual materials may struggle when those materials appear together. The distinction between adding material information to input features and using it to guide routing is useful and provides a clear motivation for the method.

The ablations help explain why the method works. Tables 2, 6, and 8 examine bounded intervention and conflict-aware gating, showing that stronger material guidance is not always better. The routing visualizations also support the intended behavior near material interfaces. These results make the design choices more convincing.

Critical weaknesses

The gains are not yet supported by strong statistical evidence. The six-seed comparison reported in Section 6 and Appendix F.2 is not significant after Holm correction (adjusted p = 0.75). This does not rule out a real improvement, but it leaves uncertainty about how consistently the method outperforms the baseline. Since these gains are central to the paper, this is a substantial concern.

The evaluation is narrow in both geometry and baselines. The experiments use simple 2D domains with circular or elliptical inclusions, and the comparisons mainly focus on Transolver variants. Results on more complex interfaces and comparisons with other architectures, such as GINO or MeshGraphNets, would help establish whether the benefits extend beyond the current setup.

Questions for authors

Which comparisons does the adjusted p-value of 0.75 refer to? Could you provide per-seed results and confidence intervals for the improvements?

Does the method remain effective on more complex material interfaces or 3D geometries, and how much overhead does computing the influence fields add?

How does CCR-MIG compare with other neural PDE solvers under the same compositional split and training budget?

2 comments
AuthorsOpenReview ↗

Response to Reviewer fDpY

Response to Reviewer fDpY (1/2)

The gains are not yet supported by strong statistical evidence. The six-seed comparison reported in Section 6 and Appendix F.2 is not significant after Holm correction (adjusted p = 0.75). This does not rule out a real improvement, but it leaves uncertainty about how consistently the method outperforms the baseline. Since these gains are central to the paper, this is a substantial concern.

We thank the Reviewer for this important point, and we agree. We rebuilt the evaluation around this concern. All models are now trained with eight seeds and evaluated on a corrected, independently generated 400-sample holdout, using exact paired sign-flip tests with Holm correction over the three primary comparisons. CCR-MIG improves on Transolver in all seven benchmarks (20–47%, 7–8 of 8 seeds, Holm p = 0.023–0.031) and on Gaussian-bank routing (40–55%, 8/8 seeds, Holm p = 0.023). It improves on the feature-matched model by 8.7–18.6% in six of these seven benchmarks (Holm p ≤ 0.039). On the rectangular thermal plate the two are on par (−4.1%, 4/8 seeds, p = 0.63), and we report this as parity. These changes are in Section 5.4 (p. 8, l. 406), Table 1 (p. 9, l. 432) and Appendix K (p. 32, l. 1703).

The evaluation is narrow in both geometry and baselines. The experiments use simple 2D domains with circular or elliptical inclusions, and the comparisons mainly focus on Transolver variants. Results on more complex interfaces and comparisons with other architectures, such as GINO or MeshGraphNets, would help establish whether the benefits extend beyond the current setup.

Thank you for this suggestion. We added waisted plates with zero to two voids, a perforated plate with a large irregular void on a non-conforming mesh, and two nonlinear constituent laws. We also added eight external operators trained with the same data, split, inputs, loss and budget: FNO, GINO, GNOT, MeshGraphNet, the Galerkin Transformer, DeepONet, PointNet and an MLP. The geometries are described in Section 5.1 and Appendix J (p. 29, l. 1530), and the external results are in Table 2 (p. 9, l. 445). Training inclusions remain circular. On 180 plates with ellipse, kidney and blob inclusions never seen in training, CCR-MIG is 21.0% more accurate than Transolver (153 of 180 plates; one training seed; p. 37, l. 1972).

Which comparisons does the adjusted p-value of 0.75 refer to? Could you provide per-seed results and confidence intervals for the improvements?

We thank the Reviewer for this question, which the original text did not answer clearly. The value referred to the four primary comparisons of the submission: CCR-MIG versus Transolver and versus the feature-matched model, for both PDEs, on the test set that we have now withdrawn. Per-seed errors for every model are listed in Tables 5–12, 13 and 14 (Appendix K, p. 30, l. 1598) and plotted in Figure 10. The 95% confidence intervals of the per-seed relative error reduction are:

Benchmarkvs. Transolvervs. Feature-matchedvs. Gaussian-bank
Elasticity, rectangle27.1% [14.5, 39.8]8.4% [3.0, 13.9]45.7% [28.7, 62.8]
Thermal, rectangle46.4% [38.3, 54.6]−4.6% [−21.8, 12.6]52.0% [37.7, 66.3]
Hyperelasticity (Fung)19.2% [10.2, 28.2]11.5% [1.2, 21.8]38.1% [25.4, 50.9]
Elasticity, waisted19.3% [10.8, 27.8]8.7% [1.9, 15.5]35.0% [21.3, 48.7]
Thermal, waisted41.4% [29.5, 53.3]9.4% [3.8, 15.0]50.8% [38.6, 63.0]
Thermal, perforated41.1% [30.5, 51.7]16.7% [4.1, 29.3]48.8% [42.9, 54.7]
Nonlinear thermal, perforated37.0% [29.8, 44.3]17.1% [3.2, 31.0]40.7% [34.1, 47.3]

All intervals exclude zero except feature-matched on the rectangular thermal plate.

Does the method remain effective on more complex material interfaces or 3D geometries, and how much overhead does computing the influence fields add?

Thank you for these questions. On plates with voids, CCR-MIG remains 20.0% (waisted elasticity), 43.5% (waisted thermal) and 42.4% (perforated thermal) below Transolver. We did not evaluate 3D geometries, and we list this as future work (p. 10, l. 526). The influence fields and descriptors are closed-form and computed once per mesh, in 1–20 ms on a CPU for about 2,800 nodes. CCR-MIG adds 6% parameters. Training takes 13.9 s versus 5.9 s per epoch (2.4×), and inference is 2.6× slower than Transolver (p. 23, l. 1188).

AuthorsOpenReview ↗

Response to Reviewer fDpY

Response to Reviewer fDpY (2/2)

How does CCR-MIG compare with other neural PDE solvers under the same compositional split and training budget?

We appreciate this question. The external operators use the same split, inputs, loss, optimizer and 200-epoch budget (Appendix H, p. 25, l. 1300). CCR-MIG is not the most accurate model on every benchmark:

mixed-400 error (CCR-MIG seed wins)CCR-MIGFNOGINO
Elasticity, rectangle0.02320.0202 (0/8)0.0211 (0/8)
Hyperelasticity (Fung), rectangle0.01860.0159 (0/8)0.0147 (0/8)
Elasticity, waisted with voids0.04840.0561 (8/8)0.0515 (6/8)
Thermal, waisted with voids0.01220.0129 (6/8)0.0128 (6/8)
Thermal, perforated0.00810.0157 (8/8)0.0071 (2/8)
Nonlinear thermal, perforated0.00940.0171 (8/8)0.0078 (0/8)

On the rectangle, CCR-MIG is on par with GNOT (5/8 seeds) and more accurate than the Galerkin Transformer, DeepONet, PointNet, MeshGraphNet and the MLP (8/8 seeds each). On waisted thermal plates, the differences from FNO and GINO are not statistically significant (unadjusted p = 0.18 and 0.27). There, GNOT reaches 0.0120 versus 0.0122 for CCR-MIG, and CCR-MIG has the lower error in three of eight seeds. We discuss why FNO and GINO are more accurate on the rectangle at p. 30, l. 1612.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers: Summary of Revisions

We thank the Reviewers for reading the manuscript and for their detailed comments. Our responses and the revisions of the manuscript are given below. In the revised manuscript, new and rewritten text is shown in blue. Locations are given as page (p.) and line (l.) of the revised PDF.

Correction of the test data. While rebuilding the evaluation we found that the mixed-composition test set used in the submission differed from the training data in its inclusion-radius and clearance ranges and in one feature convention. Part of the reported gains reflected robustness to this mismatch. We therefore withdraw the submitted numbers (17.4%, 33.5%, six seeds). Every result in the revision, for every model, uses regenerated mixed sets drawn from the training distribution.

Revised routing. The submitted model mixed one fixed Gaussian band over the slice index per material category. In the revised CCR-MIG, each constituent re-weights the node's own learned routing through a network shared by all constituents, under the same bounded blend. The submitted design is retained as the baseline "Gaussian-bank CCR-MIG".

Results. The table below summarizes the mixed-400 holdout (eight training seeds, exact paired sign-flip tests, Holm correction over the three primary comparisons).

Relative error reduction of CCR-MIG on mixed-400 (seed wins; Holm-adjusted p). Negative values mean CCR-MIG is less accurate.

BenchmarkErrorvs. Transolvervs. Feature-matchedvs. Gaussian-bank
Elasticity, rectangle0.023229.4% (7/8; .031)8.7% (7/8; .031)50.8% (8/8; .023)
Thermal, rectangle0.009846.8% (8/8; .023)−4.1% (4/8; .63)54.7% (8/8; .023)
Hyperelasticity (Fung)0.018620.2% (8/8; .023)12.0% (6/8; .039)40.4% (8/8; .023)
Elasticity, waisted0.048420.0% (7/8; .031)8.8% (7/8; .031)41.6% (8/8; .023)
Thermal, waisted0.012243.5% (8/8; .023)9.8% (7/8; .023)53.8% (8/8; .023)
Thermal, perforated0.008142.4% (8/8; .023)17.6% (6/8; .031)49.0% (8/8; .023)
Nonlinear thermal, perforated0.009438.6% (8/8; .023)18.6% (7/8; .031)41.1% (8/8; .023)

List of changes in the revised manuscript

ChangeReviewerLocation
Abstract and contributions rewrittenAllp. 1, l. 26; p. 2, l. 101
Figure 1 redrawn with rectangular, waisted and perforated platesfDpY, 8qbN, bhawp. 4, l. 173
Figure 2 redrawn: block diagram and learned routing on a mixed platebhawp. 5, l. 235
Revised routing with constituent proposals; Proposition 1bhawp. 6, l. 271; p. 7, l. 343
Protocol: 378/42 split, development set, independent 400-sample holdout; Figure 3fDpY, 8qbN, bhawp. 7, l. 370; p. 7, l. 333
Eight external neural operators under the same protocolfDpY, 8qbNp. 8, l. 378; p. 9, l. 465; Table 2 (p. 9, l. 445)
Eight seeds, exact paired tests, Holm correctionfDpY, 8qbN, bhawp. 8, l. 406; Table 1 (p. 9, l. 432)
When routing helps beyond input conditioningbhawp. 8, l. 425
Nonlinear laws: Fung hyperelasticity and temperature-dependent conduction8qbN, bhawp. 9, l. 477; p. 18, l. 962; p. 19, l. 1013
Conclusion and future work rewrittenAllp. 10, l. 508
Measured computational overheadfDpYp. 23, l. 1188
Comparison of Gaussian-bank routing and CCR-MIG (Table 3)bhawp. 23, l. 1225
Non-Gaussian (tanh) proposal controlbhawp. 24, l. 1264
Constituent interaction analysisbhawp. 24, l. 1290
Routing under input-node thinning (Figure 8)bhawp. 27, l. 1404
Per-seed errors for all models (Tables 5–12, 13, 14; Figure 10)fDpY, 8qbNp. 30, l. 1598
Why FNO and GINO are more accurate on the rectanglefDpY, 8qbNp. 30, l. 1612
Perforated-plate benchmark (Figure 11, Table 13)fDpY, 8qbN, bhawp. 31, l. 1661
Boundary conditions and example solutions for the perforated and waisted plates (Figures 6 and 9)fDpY, 8qbN, bhawAppendices C.2 and J.6
In-distribution accuracy8qbNp. 32, l. 1718
Conflict-gate ablation (Table 16, Figure 15)8qbN, bhawp. 37, l. 1996
Mixed-trained reference model (Table 15)8qbNp. 33, l. 1757
Qualitative zero-shot predictions (Figure 4 and Figure 12)bhawp. 10, l. 501
Test on non-circular inclusions (Figure 13)fDpY, 8qbN, bhawp. 37, l. 1972

Limitations stated in the revision. All benchmarks are two-dimensional, and training uses circular inclusions only; the test on non-circular inclusions uses one training seed. Routing is on par with feature conditioning on the rectangular thermal plate. FNO and GINO are more accurate than CCR-MIG on the void-free rectangle, and GINO is also more accurate on the perforated plate. The temperature-dependent conduction benchmark is only weakly nonlinear (median deviation from linear conduction 0.09%).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.