acceptodds
Under review as a conference paper at ICLR 2027

CoRaWS: Zero-Shot Neural Warm Starts across Topology Families for Nonlinear Magnetostatics

Abstract

Neural network surrogates have attracted broad interest because they can support repeated physical simulations at low inference cost. However, if deployment on an unseen target topology family still requires collecting high-fidelity labels and training a model, the preparation cost may offset the gains at inference. Even when field prediction errors are low, the predicted solution may not satisfy the numerical solver's convergence criteria. Applications to unseen topology families therefore need to control data preparation costs while ensuring the accuracy of the final solution. For nonlinear magnetostatics, we propose complementary approaches to data construction and initial-guess prediction. We first develop EMTopoGen, a parametric topology generator combining primitive shapes, assembly relations, and continuous parameters. Using it, we construct EMTopoBench-54 with 54 topology families and 27,000 high-fidelity simulation samples. This benchmark supports the study of compositional generalization through training on source families and evaluation on unseen target families. We then develop CoRaWS, a neural warm-start method based on constitutive coordinate reparameterization and field-magnitude awareness. After training on source families, the method generates finite element initial guesses for unseen target topology families without target-family labels or fine-tuning, accelerating the solution process while maintaining high fidelity. Experiments on nonlinear magnetostatic problems show that CoRaWS offers advantages for FEM acceleration, indicating its potential for engineering simulation. Code will be released upon acceptance.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Scientific Machine Learning, Physics-Informed Machine Learning, Neural Operators, Numerical Methods for PDEs, Computational Electromagnetics

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_R2hwOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present EMTopoBench-54, a nonlinear magnetostatics benchmark with various topology families and 500 samples per family. They also develop CoRaWS, an architecture for neural warm-starts of numerical solvers. They train on source families of their benchmark and transfer to unseen target topologies, while still showing an acceleration of the solving time.

Critical strengths

This paper introduces an interesting new benchmark to the machine learning community, which is itself valuable. The authors add that this benchmark is intended for evaluating the transferability of methods to unseen topologies at test time. The experimental results and ablations are thorough and demonstrate that CoRaWS is still capable of accelerating solving time even at topologies unseen during training.

Critical weaknesses

While the goal of introducing a new scientific benchmark is valuable to the machine learning community, a few additions to this paper would strengthen this. Currently, the benchmark is framed in the abstract as supporting “the study of compositional generalization through training on source families and evaluation on unseen target families.” A scientific benchmark with this purpose would be very valuable to the community. However, the focus of the paper (correct me if I am misunderstanding) is specifically for neural warm-start methods. As such, to present a more general benchmark, I would recommend including comparisons of existing trained surrogate models (no need to develop new ones here, as CoRaWS already does this for the warm-start setting), i.e., models without solvers-in-the-loop. This comparison would evaluate transferability to unseen topologies under the common surrogate modeling setting, and would strengthen the position of the benchmark substantially. The other option would be to constrain the settings where the benchmark should apply, specifically to neural warm-start methods. However, this runs the risk of limiting the reach and impact of the benchmark.

On a more minor note: Table 1 measures wall-clock time, but do you have any indication of how CoRaWS compares to other methods in terms of FLOPs and other timing metrics? I understand it may be computationally infeasible to recompute the precise FLOP count, but a rough estimate would be valuable because wall-clock comparisons could differ across machines.

While the paper is well-written, I believe there are some sections where machine learning audiences may struggle to understand. Concretely, I would recommend:

  • Emphasizing why this benchmark is important and what it adds over previous benchmarks
  • Modifying some of the background sections in the main text. Some sections (e.g., eq 1) are too dense and might benefit from simplifications or expanded explanations. For instance, if this background is just to determine that “We therefore predict HH in magnetic materials and recover BB through the inverse constitutive relation.”, then perhaps the background can be moved into the appendix and this result can be emphasized in the main text.

I am open to improving my score.

Questions for authors

Do multiple resolutions matter in this application? If so, using ConvNeXt style network (i.e., not operator) architectures may limit generalization to unseen resolutions. This may be important if the benchmark considered resolutions unseen during training, but I am not very familiar with the application domain.

6 comments
AuthorsOpenReview ↗

Response to Reviewer R2hw — Part 1/4

Dear Reviewer R2hw,

Thank you for suggesting that we demonstrate the benchmark's use in conventional surrogate modeling, and for the encouragement of your willingness to consider raising your score; we will carefully follow your suggestions to further improve the paper. We have added cross-topology-family direct prediction with FNO and GAOT and stratification by field magnitude using a common criterion, quantified network computation and numerical solution work, and tested unseen input resolutions. We will use these additions to strengthen the benchmark introduction and present the physical intuition and evaluation objectives before the mathematical details. We have completed the relevant revisions in response to your suggestions, and we hope that these changes and additional evidence fully address your concerns and encourage you to reconsider your assessment of our work. If any issues still affect your judgment, we would be happy to clarify them further and make additional improvements.

W1 | The benchmark's use with conventional surrogate models

Thank you for suggesting that we validate the benchmark's uses separately from CoRaWS itself. We have added conventional field-prediction experiments with FNO and the geometry-aware operator GAOT, directly evaluating their transfer to unseen topology families without a numerical solver during prediction. The geometries, materials, excitations, field labels, and topology-family splits provided by EMTopoBench-54 support this evaluation. We have added a dedicated section to introduce this, and it does have potential for application to other domains.

All models use the same three splits, with 12,000 source-family training samples, 120 epochs, and 480 target-family pairs per split. FNO and GAOT select between linear and radial asinh output representations on 600 held-out source-family samples, training both candidates for 12 epochs and continuing the selected candidate to 120 epochs, with training of the unselected candidate counted separately as selection cost.

ModelFold 0Fold 1Fold 2All pairs
Direct-B68.0361.5449.6658.35
CoRaWS, H recovery76.6780.6945.6167.82
FNO80.0980.7355.8073.40
GAOT93.51105.4588.1996.07

To better understand which samples account for these errors, we stratify by ground-truth field RMS within the same valid region V\mathcal V:

Brms∗=1∣V∣∑x∈V∥B∗(x)∥22.B_{\mathrm{rms}}^{*}=\sqrt{\frac{1}{|\mathcal V|}\sum_{x\in\mathcal V}\|\mathbf B^{*}(x)\|_2^2}.

With thresholds of 1 and 100 mT, all models use exactly the same samples.

Ground-truth field Brms∗B_{\mathrm{rms}}^{*} (mT)PairsDirect-BCoRaWS, H recoveryFNOGAOT
<1<146463.3377.1076.83107.53
[1,100)[1,100)70063.9273.4277.7496.64
≥100\ge10027626.9916.6327.2560.33

The stratified results explain why the overall median is high: 1,164 pairs have a ground-truth field RMS below 100 mT, accounting for 80.83% of all pairs. These samples are numerous and have high relative errors, so the overall median primarily reflects prediction performance at lower field magnitudes. By contrast, CoRaWS has a median error of 16.63% in the group with a ground-truth field RMS of at least 100 mT. In fact, relative error uses the ground-truth field norm as its denominator, so small absolute deviations in weak fields can correspond to large relative errors. For example, in the group below 1 mT, CoRaWS's median absolute field-error RMS is approximately 0.392 mT. Although small in absolute terms, this deviation is not negligible relative to the actual field.

Following your suggestion, we have added direct evidence of the benchmark's use in conventional surrogate modeling, revealing prediction challenges at different field magnitudes. Direct-B has lower errors in the two weak-field groups, while CoRaWS's H recovery achieves a lower error in the group at or above 100 mT. This helps explain why routing in the original manuscript can more effectively reduce subsequent solver work in warm starts. It also allows the same benchmark to evaluate both field prediction and the assistance predictions provide to numerical solution. We will therefore retain EMTopoBench-54's focus on compositional generalization, add conventional surrogate-model comparisons in Section 4.2, and report the common field-magnitude stratification and training and evaluation settings in Appendix E.4.

AuthorsOpenReview ↗

Response to Reviewer R2hw — Part 2/4

W2 | Network computation and numerical solver workload

Thank you for suggesting additional computational metrics. We quantify network forward computation and numerical solver work separately to explain the sources of timing differences. At a 256×256256\times256 input resolution, CoRaWS's forward pass includes three-head prediction, magnitude classification, and H-to-B recovery. The PyTorch profiler counts are as follows:

ModelParameters (M)Operations per forward pass (GFLOPs, partial count)
Direct-B30.67826.050
CoRaWS30.77926.054

We additionally calculate normalization and bilinear interpolation costs from the actual tensor dimensions and estimate constitutive recovery costs using the 64-point material table. One multiply-add is counted as two floating-point operations.

OperationDirect-B (GFLOPs)CoRaWS (GFLOPs)
LayerNorm / GroupNormApproximately 0.045Approximately 0.045
Bilinear interpolation, including coordinate and weight calculationsApproximately 0.023–0.046Approximately 0.026–0.050
Magnitude interpolation and vector recovery after constitutive lookup—Approximately 0.00059

Normalization is estimated at approximately 8 basic operations per element, and the interpolation range reflects whether coordinates and weights are reused across channels. For 2562=65,536256^2=65{,}536 raster points, constitutive recovery additionally involves 65,536 square roots and at most 458,752 binary-search comparisons, with indexing, comparisons, and memory accesses not counted as multiply-adds. Interpolation here refers to upsampling within the network, while field transfer to FEM quadrature points is counted separately as initial-guess construction. After deducting the constitutive multiply-adds already counted, the supplemented arithmetic counts are approximately 26.12–26.14 GFLOPs for Direct-B and 26.13–26.15 GFLOPs for CoRaWS. Most computation still comes from the shared backbone. Special functions such as exponentials and logarithms cannot be interpreted in terms of ordinary multiply-add time, so the complete cost is still measured by online time.

Sparse FEM does not lend itself to a single FLOP count per update because the degrees of freedom, matrix structure, and factorization fill-in vary across samples, and line search requires repeated computation. We therefore separately count coarse- and fine-mesh updates and initial-guess constructions across all 1,440 pairs:

MethodFine-mesh Newton updatesCoarse-mesh Newton updatesFull network forward passesB→Az projections
Zero initialization7,662000
Direct-B7,19301,4401,440
CoRaWS, routing off6,77301,4401,440
CoRaWS, routing on6,75901,440319
C2F5,59715,70600
C2F, same qHq_H routing5,80512,8821,4400

CoRaWS reduces fine-mesh updates by 11.79% relative to zero initialization. Routing still performs 1,440 full forward passes but reduces projections from 1,440 to 319, avoiding construction overhead on easy samples. C2F instead trades coarse-mesh solution and transfer for fewer fine-mesh updates. C2F without routing includes 2,122 coarse-mesh refinements, compared with 957 under the same routing. These metrics complement wall-clock time by showing how computational work is reduced or shifted, and will be reported in Appendix E.5.

AuthorsOpenReview ↗

Response to Reviewer R2hw — Part 3/4

W3 | Presenting the benchmark's value and the method's intuition

Following your suggestion, we will present the benchmark's value and the method's intuition before the details. The benchmark section will first explain the evaluation capabilities it adds, and the method section will explain why we predict H before developing the structural definitions and mathematical analysis. Readers can thus first understand the problem and design objectives, with detailed derivations available in the appendix.

The benchmark's main value is to explicitly distinguish parameter changes from changes in assembly structure. Varying dimensions or excitation while keeping component combinations and assembly relations fixed presents the model with parameter changes within the same structure. Changing component combinations and assembly relations, however, can alter magnetic circuit connections and branches, requiring the model to transfer to new structures. EMTopoGen explicitly defines primitives, assembly relations, and continuous parameters, on which EMTopoBench-54 builds 54 topology families and 27,000 samples and evaluates structural transfer by holding out entire families.

We will emphasize this evaluation dimension in the introduction and benchmark description and use the new FNO and GAOT results to show that the same benchmark can evaluate both field prediction on unknown structures and whether those predictions assist numerical solution. The structural grouping in the appendix further distinguishes unseen families within seen engineering groups, unseen engineering groups composed of seen primitives, and target families containing unseen primitives, allowing readers to see the difficulty and outcomes of each type of transfer.

For the method derivation, Section 3.2 will first explain the intuition: at different magnetization states, the same B deviation can correspond to different H deviations, so the material constitutive relation provides a physically motivated way to measure error. We will then explain why we predict H in magnetic materials and recover B through the known constitutive relation. The scalar expansion and detailed derivations are moved to Appendix C.1, while the main text retains the vector form and its connection to the FEM tangent response. Section 3.3 will likewise explain the roles of encoding, magnitude classification, and the prediction heads before introducing the equations.

We believe these changes preserve the theoretical basis of the method while allowing readers to first understand its design objectives. The local analysis explains how the representation changes error weighting, while experiments test whether it reduces solver work.

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

This paper presents a framework for accelerating nonlinear magnetostatic solves, including a parametric topology generator (EMTopoGen) used to construct a benchmark of 54 topology families and 27,000 high-fidelity simulation samples, and a neural warm-start method (CoRaWS). At inference time, the model generates initial guesses for the unseen target topology families without requiring any labels or fine-tuning. The zero-shot experiments show that both the required number of Newton iterations and the cumulative online time can be reduced.

Critical strengths
  1. This paper is generally well written, and the motivation for each technical component is clearly explained. This paper pinpoints a practical problem: beyond fast inference and accurate final predictions, it is also important to control the data preparation costs and reduce the total solution time, while ensuring that the predicted solution satisfies the numerical solver’s convergence criteria.

  2. The experiments and ablation study (Table 2 and Figure 4) are thorough. The experiments in Section 4.3 help verify the contributions of constitutive recovery and initialization routing, and also show that routing is mainly to avoid construction overhead for samples that are easy to solve from zero.

Critical weaknesses
  1. Presentation needs improvement.
  • The first paragraph of the introduction section includes too many details of the existing methods. Some of these details could be moved to the related work section.
  • Section 3.3 contains dense information and complex equations. Providing some intuition first would make it easier to follow.
  1. Concerns regarding method and results.
  • Section 2.2 reviewed several methods on warm-start, such as Learning to Warm-Start, Neural-Initialized Newton, and PFEM, but the experiments mainly compare to specific initialization strategies not to the full approach (Section 4.1).
  • In Section 4.2, the authors report that CoRaWS reduced cumulative online time by 8.41%, including network inference and initial-guess construction, but this overall saving seems modest. The authors could further discuss whether this reduction is significant.
Questions for authors
  1. The authors have acknowledged that the evaluation is limited to magnetostatics, but the paper should explain in more details on why this problem was chosen. Moreover, is there any part of this framework designed specifically for nonlinear magnetostatics, and what would need to be adapted for applications in other domains?

  2. Could the authors justify the choices of baseline approaches and clarify whether the methods discussed in Section 2.2 can be directly compared?

7 comments
AuthorsOpenReview ↗

Response to Reviewer SPSB — Part 1/3

Dear Reviewer SPSB,

Thank you for recognizing the practical significance of the problem and for your suggestions on baseline comparisons and presentation. We have added magnetostatic adaptations of NiN, PFEM, and Learning to Warm-Start, and used shared routing to compare the contributions of the initial guesses themselves. We will also explain the practical significance of acceleration in the context of repeated solves and cross-family reuse, rewrite the introduction to EMTopoGen, and present the physical intuition before the method's equations. We have completed the relevant revisions in response to your suggestions, and we hope that these changes and additional evidence fully address your concerns and encourage you to reconsider your assessment of our work. If any issues still affect your judgment, we would be happy to clarify them further and make additional improvements.

W1 | Readability of the introduction and method

In the introduction, we will retain the connection between fast field prediction, numerical convergence requirements, and the cost of preparing data for new topology families, while moving implementation details of existing methods to the related work so that readers first see the problem this paper addresses.

For Section 3.3, we will first explain the purpose of each component: invertible encoding handles the wide dynamic range of field values, joint supervision helps the model learn field distributions and material responses, and magnitude prediction also determines whether constructing a neural initial guess is worthwhile. We will introduce the equations after establishing these relationships. The scalar error derivation will be moved to Appendix C.1, while the main text retains the physical explanation and its connection to vector error weighting, allowing readers to understand the mathematical formulation through the design objectives.

W2 | Complete learning frameworks and the practical significance of acceleration

We have added NiN, PFEM, and Learning to Warm-Start. All three methods have completed evaluation on all 1,440 pairs, with additional shared-routing comparisons for NiN and PFEM.

NiN trains a modulated implicit network using a physics-based objective and performs three latent-variable updates during inference, which are included in the online cost. PFEM trains its predictor using magnetic potential energy under finite-element discretization. Each fold uses the same 12,000 source-family training samples and a 120-epoch budget, without training labels from target families. Final evaluation uses a common fine mesh, Newton solver, and acceptance criteria, and reuses the original CoRaWS's qHq_H decisions when routing is enabled.

MethodRoutingEvaluated pairsTotal Newton updatesCumulative online time (h)
Zero initializationNone1,4407,66219.4688
NiNOff1,4408,14919.7509
NiNOn1,4407,95719.4659
PFEMOff1,44016,37334.0651
PFEMOn1,4409,67722.4249
Learning to Warm-StartOff1,44028,53755.1573
CoRaWSOff1,4406,77319.8110
CoRaWSOn1,4406,75917.8321

Learning to Warm-Start requires a total of 28,537 Newton updates and 55.1573 h across all 1,440 pairs, of which 1,369 pass final acceptance. Under this training budget, training through unrolled iterations does not yield net acceleration. Shared routing reduces the cumulative online time of both NiN and PFEM. However, NiN's cumulative time is almost identical to that of zero initialization, with a difference of only approximately 0.015% that is insufficient to establish a reliable net acceleration, while PFEM remains 15.18% slower than zero initialization. Under the same routing, CoRaWS still requires fewer Newton updates and less total time than NiN and PFEM, showing that this difference cannot be explained by the selection mechanism alone.

The 8.41% online saving can accumulate over repeated solves. For example, device parameter optimization often requires repeated FEM solves for different parameter configurations of the same topology. The offline investment still needs to be amortized. If the system is built from scratch for only a small number of solves, CoRaWS offers no computational cost advantage. Only in scenarios involving sustained repeated solves, such as large-scale parameter optimization, can cross-family reuse allow cumulative savings to offset the initial investment. Appendix E.5 reports estimated label generation costs, training costs, and the corresponding query counts needed for amortization, distinguishing deployment gains when source data and weights already exist from the investment required to build the entire system from scratch.

AuthorsOpenReview ↗

Response to Reviewer SPSB — Part 2/3

Q1 | The engineering significance of magnetostatics and the general design of EMTopoGen

The revised manuscript adds an explanation of EMTopoGen's general design and its relationship to magnetostatic applications. EMTopoGen separates compositional geometry generation from physical task configuration: primitive shapes, assembly relations, and parameter dependencies jointly define a geometry template, sampling free parameters produces individual geometries, and objects, hierarchies, and geometric relations are then checked for validity. This process does not depend on a B–H curve, magnetic field variables, or a particular solver, and can independently generate and validate two-dimensional geometries.

Only after geometry generation does the downstream task assign material properties to regions, specify excitations, boundary conditions, and output quantities, and pass them to the corresponding solver. Thus, EMTopoGen's mechanisms for geometry composition, parameter dependencies, and sampling can be reused for other physical tasks, with the required material laws, loads, boundary conditions, and solver interfaces configured for the target problem. EMTopoBench-54 is a specific instance of this design in nonlinear magnetostatics, rather than the full scope of EMTopoGen itself.

We chose nonlinear magnetostatics primarily because it directly supports the design and optimization of electric machines, transformers, and magnetic devices. Engineering design requires repeated comparisons of component dimensions, assembly relations, and excitation conditions, making reliable magnetic field solution both the basis of performance evaluation and a computational burden in large-scale parameter searches. Reducing the time needed to obtain a converged solution therefore has direct engineering value. At the same time, magnetostatics shares a similar modeling and solution process with common FEM tasks such as structural mechanics and heat transfer: materials, loads, and boundary conditions are defined on a given geometry, finite-element discretization produces algebraic equations, and numerical solution yields the physical field. For problems such as nonlinear elasticity and temperature-dependent heat conduction, material responses also vary with the current state and require iterative methods such as Newton's method. These commonalities make the idea of organizing prediction representations through material constitutive relations and then using the original solver for numerical correction worthy of broader investigation. This work implements that idea through the B–H relation, an auxiliary permeability head, and magnetic vector potential projection, while extensions to other problems can construct prediction representations and initial-guess interfaces around the corresponding constitutive relations and solution variables.

We will provide a complete introduction to EMTopoGen in a dedicated subsection of the main text, rewriting the current brief Section 3.1 to explain in sequence how compositional geometries are generated, how physical tasks are configured independently, how continuous geometries are converted into a common raster representation, and how EMTopoBench-54 is constructed on this basis. The common representation preserves the physical dimensions and relative positions of components while separating the network input from the specific FEM mesh. Appendix C.3 adds details of sampling and coordinate normalization, while the conclusion notes that adaptation across physical tasks still requires validation.

AuthorsOpenReview ↗

Response to Reviewer SPSB — Part 3/3

Q2 | Baseline selection and comparison of complete methods

The original baselines balance representativeness in neural warm starts with the competitiveness of numerical methods used in engineering. Direct-B represents the typical approach of directly predicting a physical field and then correcting it with a solver, constructing the initial guess through a common projection. C2F represents an established coarse-to-fine solution strategy that first solves on a coarse mesh and then initializes the fine mesh, without requiring training labels or a network. Thus, the two respectively test the benefit of constitutive-aware prediction over direct prediction and the value of learned initial guesses relative to traditional numerical initialization.

BaselineWhat it tests
Direct-BDirect field prediction followed by the common projection to obtain an initial guess
Direct-AzDirect prediction of the potential variable used by the solver
POD-ridge-AzGeneration of initial guesses through a reduced-order representation and regression
Nearest neighborReuse of available solved examples
C2FInitialization of the fine mesh using a numerical coarse-mesh solution
Linear initialization with initial permeabilityUse of the full linear solution obtained from the material's zero-field linearization

These strategies share the final numerical correction process, while the B/H ablation with the same weights further fixes the model, routing, and projection. The additional NiN, PFEM, and Learning to Warm-Start comparisons examine different training procedures, incorporating physics-based latent-variable updates and meta-training, finite-element physics-based training, and iteration unrolling, respectively, into the material relations, boundaries, and discrete operators of this problem. The full-evaluation times without routing for NiN, PFEM, and Learning to Warm-Start are 19.7509 h, 34.0651 h, and 55.1573 h, respectively, all higher than with zero initialization.

Reviewer_TB3DOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies neural warm starts for 2D nonlinear magnetostatic FEM solves on assembly structures unseen in training. It introduces a parametric generator (EMTopoGen) and a benchmark (EMTopoBench-54) of 54 topology families with 500 Maxwell-labeled samples each, evaluated by holding out entire families. The proposed method, CoRaWS, predicts HH inside magnetic material and recovers BB through the known BB–HH curve, uses a nonlinear magnitude encoding, supervises BB, HH, and μ\mu heads jointly, and uses a magnitude classifier to route each sample to either zero or neural initialization. On 1,440 target pairs, with matched convergence criteria and all online costs counted, the authors report reductions of 8.41% in cumulative online time and 11.79% in Newton iterations relative to zero initialization.

Critical strengths
  1. The evaluation seems careful: Speedups are measured end to end at matched accuracy, with initialization overhead included, and the paper shows that naive learned warm starts can be slower than zero initialization.

  2. The idea of predicting HH and recovering BB through the material law appears well motivated by the link between HH-space error and the reluctivities in the Newton tangent matrix. A controlled ablation seems to support it: with weights and routing fixed, switching only the recovery path reduces Newton updates from 7,219 to 6,759.

  3. The authors are open about the Fold 1 slowdown and the use of F32 during development, and the family-holdout benchmark could be a useful resource if released.

Critical weaknesses
  1. A large part of the net gain appears to come from routing, which does not seem to be applied to the baselines. Without routing, CoRaWS is reported to be slightly slower than zero initialization (19.81 h vs 19.47 h). Since C2F already improves on zero initialization without routing (18.67 h), it might be informative to see whether it would remain competitive with the same routing. Given the reported 0.986 AUROC, a simpler non-learned router might also work reasonably well.

  2. The comparison might benefit from standard physics-based warm starts, such as initializing from the linear solve with initial permeability. Some learned baselines may not be fully tuned (POD-ridge shows about 7,850% initial error), and including one of the cited geometry-aware neural operators could strengthen the comparison.

  3. The absolute benefit seems relatively modest, at roughly 4 seconds per solve, and a break-even analysis against the offline cost (about 12,000 Maxwell solves plus training per split) would help put it in context. The cumulative metric may also obscure how gains are distributed. If my reading of Table 4 is correct, CoRaWS is about 20% faster on the 281 high-iteration pairs but possibly around 10% slower on the remaining 1,159.

  4. Variance across seeds or repeated timings does not seem to be reported, which makes some small ablation differences (e.g., 4.69 vs 4.70 iterations) somewhat hard to interpret. A breakdown by fully held-out versus seen engineering groups might help support the compositional generalization claim. The reported 100% initial error for CoRaWS appears to reflect routing rather than prediction quality, so the error on neurally initialized pairs might be more informative.

  5. The methodological novelty may be somewhat limited. The main contribution appears to be a choice of output coordinate supported by a first-order analysis, and the class-conditioned encoding does not seem to yield a measurable benefit. Since the experiments use a single isotropic BB–HH curve in 2D, it is not yet clear whether the principle would extend to other materials or physics.

Questions for authors
  1. Would Table 1 change much if the same routing were applied to Direct-B and C2F, and how might a non-learned router compare?
  2. How would CoRaWS compare to initializing from the linear solve with initial permeability?
  3. Would it be possible to report results separately for fully held-out and seen engineering groups, along with per-sample speedup distributions?
  4. Is there any information on variance across training seeds, or on initial error for neurally initialized pairs?
  5. Roughly how many queries might be needed to amortize the offline cost?
7 comments
AuthorsOpenReview ↗

Response to Reviewer TB3D — Part 1/5

Dear Reviewer TB3D,

Thank you for your careful assessment of routing fairness, constitutive representation, and practical benefits. We have added comparisons using shared routing, non-learned routing, linear initialization with initial permeability, POD parameter selection, and geometry-aware models, and examined training seeds, the distribution of gains, and other material curves. The new results confirm that routing also strengthens numerical baselines without surpassing the enhanced CoRaWS, while comparisons with the same weights and routing still show that H recovery reduces Newton updates. We have completed the relevant revisions in response to your suggestions, and we hope that these changes and additional evidence fully address your concerns and encourage you to reconsider your assessment of our work. If any issues still affect your judgment, we would be happy to clarify them further and make additional improvements.

W1 | The respective contributions of routing and constitutive recovery

You raise a highly informed point, and the shared-routing comparisons show that routing can also reduce baseline costs. Applying the original selector to C2F and including the full forward-pass cost reduces its total time from 18.6659 h to 17.0815 h. A non-learned two-step Newton probe followed by C2F takes 17.3691 h.

Meanwhile, with the model architecture, data, and 120-epoch budget unchanged, using only a cosine learning-rate schedule reduces CoRaWS's total time from 17.8321 h to 16.9118 h, which is 0.99% lower than C2F with shared routing. This improvement shows that CoRaWS can be combined with general training techniques and that the original configuration still has room for improvement.

The importance of routing does not rule out the contribution of constitutive recovery. With the same checkpoint, routing decisions, and projection, the B head requires 7,219 Newton updates, whereas H recovery requires 6,759. The two paths select exactly the same samples, so this improvement comes from how the initial guess is constructed rather than from selecting different samples.

We have also added shared-routing comparisons for the newly included NiN and PFEM:

MethodRoutingTotal Newton updatesCumulative online time (h)
NiNOff8,14919.7509
NiNSame qHq_H7,95719.4659
PFEMOff16,37334.0651
PFEMSame qHq_H9,67722.4249

Shared routing improves both methods, but NiN's time is essentially the same as the 19.4688 h of zero initialization, and PFEM remains 15.18% slower. Neither matches the original CoRaWS's 17.8321 h. We will retain the main comparison with the original model, add these comparisons in Appendices E.1 and E.3, and explain the respective roles of routing and constitutive recovery in Section 4.2.

AuthorsOpenReview ↗

Response to Reviewer TB3D — Part 2/5

W2 | Physics-based initialization, POD tuning, and geometry-aware baselines

We have added linear initialization with initial permeability, POD with parameters selected only on held-out source-family samples, and a GAOT comparison, but none of these baselines surpasses CoRaWS.

In fact, the original solve from zero initialization already computes the first Newton direction through the initial-permeability system and then performs a line search. The new baseline starts directly from the full linear solution to test whether accepting this state improves the subsequent solve. Across all 1,440 pairs, it requires 8,483 subsequent fine-mesh updates and 22.0906 h, including assembly and solution of the preliminary linear system. Zero initialization requires 7,662 updates and 19.4688 h. Thus, directly accepting the full linear solution does not outperform the solve from zero initialization with line search.

POD parameter selection uses 12,000 source-family training samples and 600 held-out source-family samples per split, searching over 4 ranks (16, 32, 64, 128) and 5 ridge regression coefficients (10−610^{-6}, 10−410^{-4}, 10−210^{-2}, 1, 100). Parameters are selected by the relative error of the recovered field, first averaged within each family and then averaged equally across families. Neither target-family labels nor solution times are used for selection, and the final FEM evaluation retains the original settings.

POD configurationMedian initial error (%)Total Newton updatesCumulative online time (h)
Original fixed configuration7,849.9318,21739.8577
Parameters selected on held-out source-family samples2,173.4810,88424.8306

Parameter selection reduces cumulative time by 37.70%. Both prediction quality and solution cost improve after tuning, but total time remains higher than with zero initialization, yielding no net acceleration.

When GAOT is used as the CoRaWS backbone, the B head and H path require 8,649 and 8,908 updates, respectively, under the same routing, and do not outperform the original backbone. We have also added conventional field-prediction experiments with FNO and GAOT without a solver. The additional numerical, reduced-order, and geometry-aware comparisons will appear in Appendices E.1, E.3, and E.4.

W3 | Average gains, offline investment, and the distribution of gains

The significance of saving approximately 4 s per solve on average lies in its accumulation over repeated solves. For example, device parameter optimization repeatedly calls FEM for different dimensions and excitations of the same topology, and searching across multiple topologies further increases the number of solves. CoRaWS's source data and pretrained weights can be reused across target families without collecting new labels or fine-tuning for each new family, so this investment is not incurred again at every structural transfer.

Building the system from scratch for only a small number of solves offers no computational cost advantage. The original training takes an average of 2.280 h per fold. At a saving of 4.092 s per query, the training cost corresponds to approximately 2,006 queries when source labels already exist. Generating 12,000 source labels takes approximately 157.27 h on average, which, together with training, corresponds to approximately 140,000 queries. Q5 gives the costs for different scopes of system construction, which will be reported together in Appendix E.5.

The stratified results show a saving of 19.71% on the 281 high-iteration pairs and an increase of 9.79% on the remaining 1,159 pairs. We consider this concentration consistent with how warm starts work: constructing an initial guess pays off only when it avoids enough nonlinear solution work. This is also why we introduced routing. We will add per-family and per-sample results in Appendix E.2 to show which solution scenarios contribute to the cumulative gains.

W4 | Training variation, repeated timing, and structural groups

We have added evidence distinguishing training variation, timing fluctuations, and differences in target structures. In fact, the main conclusion does not rest on the ablation difference between 4.69 and 4.70. The original CoRaWS requires 6,759, 6,839, and 6,855 Newton updates under three training seeds, all below the 7,662 required by zero initialization. Cosine training further reduces this to 6,492 updates. We have revised Section 4.3 and no longer describe the small difference from class conditioning as an additional reduction in iterations.

The structural grouping shows savings of 15.93% for unseen families within seen engineering groups and 6.22% for unseen engineering groups composed of seen primitives. We consider this finding related to the difficulty of generalization, which is clearly greater for unseen engineering groups composed of seen primitives.

AuthorsOpenReview ↗

Response to Reviewer TB3D — Part 3/5

W5 | Methodological contribution and applicability to other materials

CoRaWS has a relatively simple final form, but achieving zero-shot net acceleration on unseen topology families requires addressing constitutive representation, the wide dynamic range of field values, and initial-guess construction costs together.

Field magnitudes span multiple orders of magnitude as excitation varies. Fixed-range Sigmoid or tanh outputs struggle to preserve both weak-field variations and strong-field magnitudes, so we use a range-aware invertible radial encoding. Taking H as an example, magnitude labels are determined by clustering the root-mean-square field strengths of source-family training samples in logarithmic space, and the predicted probabilities adjust bounded parameters of the same invertible mapping. Magnitude classification participates in both encoding and recovery and also selects between zero and neural initialization. The existing ablations support the role of nonlinear encoding. The absence of a further reduction in iterations from class conditioning does not negate the role of magnitude classification in encoding and routing.

In fact, we also compared cross-supervision, recovered-field error, magnetic energy loss, error supervision after one Newton update, and different prediction quantities and recovery methods. The more complex designs did not consistently reduce Newton work, so we retained the simple combination of direct supervision and constitutive recovery. The contribution of this work is to connect a physically motivated prediction coordinate system to solution costs on unseen topology families, rather than to propose a new backbone. The first-order analysis explains the error weighting, and the solver experiments test the acceleration. The fact that the H path retains its gains after cosine training improves the predictor also supports the role of this representation.

We further trained separate models for other constitutive curves, using 100 training geometries, 20 evaluation geometries, the original encoding, and 120 epochs for each group. The B/H paths of the same checkpoint use the same interface interpolation and projection, without routing.

Material curveZero-initialization updatesB-head updatesH-recovery updatesH time saving relative to zero initialization
S-shaped curve6065905726.79%
VACODUR 502813172515.31%

In both groups, H recovery reduces updates relative to the B head with the same weights and saves total time relative to zero initialization. The VACODUR curve comes from the manufacturer material library provided with Ansys 2025 R1. Appendix E.6 will describe the material and interpolation settings and report the costs. These results support applicability to other constitutive relations, while zero-shot transfer between materials and extension to other physical problems remain directions for future research, as stated in the conclusion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.