acceptodds
Under review as a conference paper at ICLR 2027

GMNO on MaxwellBench: Scaling Pretrained General Mesh Neural Operator Across Industrial-Standard Electromagnetism Simulations

26 views

Abstract

Engineers routinely solve partial differential equations (PDEs) numerically to simulate physical phenomena and evaluate industrial designs. To alleviate the high computational cost of these simulations, neural operators have become a promising class of surrogate models for fast inference. While recent works have advanced neural operator architectures and begun to explore their scaling properties, most studies rely on heavily simplified PDE data that fail to reflect the structural complexity of industrial simulations. In practice, industrial simulation data have complex structure, require substantial domain expertise to interpret, and remain largely inaccessible to the broader machine learning community. We argue that a true foundation neural operator should generalize across industrial simulation scenarios that go far beyond simple parameter variations, while remaining governed by the same underlying PDEs. To bridge this gap, we introduce MaxwellBench, the first open-source industrial-standard benchmark comprising 14 datasets from distinct electromagnetism applications, and propose the General Mesh Neural Operator (GMNO), a neural operator with novel embedding architectures for handling complex multi-modal PDE data defined on meshes. We also develop a heterogeneous parallel training paradigm that enables large scale training on industrial-standard datasets. Extensive experiments show that GMNO consistently outperforms existing neural operator baselines on industrial-standard datasets. Furthermore, large scale pretraining yields strong out-of-distribution generalization, highlighting the promise of scaling neural operators on industrial applications.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operator, Pretraining, Partial Differential Equations, Out-of-Distribution Generalization, Benchmark

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_ZkQ4OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MaxwellBench, a dataset of two-dimensional Maxwell problems spanning multiple industrially relevant electromagnetic simulation scenarios, together with a General Mesh Neural Operator (GMNO). GMNO jointly encodes mesh geometry and heterogeneous physical information associated with triangular elements and line segments. The authors further investigate joint pretraining across multiple datasets, reporting improved performance over models trained from scratch on individual datasets and demonstrating few-shot transfer to related electromagnetic tasks held out from pretraining.

Critical strengths
  1. MaxwellBench provides a valuable benchmark for neural operators by covering multiple industrially relevant electromagnetic scenarios while retaining mesh structure and physical information.
  2. GMNO achieves promising predictive performance through cross-dataset pretraining and demonstrates few-shot adaptation to related electromagnetic tasks.
Critical weaknesses
  1. As a central methodological contribution of GMNO is its explicit use of mesh geometry and structure, its relationship to prior work deserves closer discussion. Earlier methods already exploit information beyond point coordinates, including element connectivity in FEN [1] and triangle-based geometric operators in PhyMPGN [2]. These works are neither cited nor discussed. Clarifying how GMNO’s mesh representation and encoding differ from these approaches would help establish its methodological novelty.
  2. The experimental evaluation raises concerns about both comparison fairness and general applicability. Although Appendix C states that boundary conditions are included in the baseline inputs, the supplementary implementations of Transolver, GAOT, and LNO do not explicitly use boundary-condition types and normals in prediction. Transolver and GAOT retain an averaged node-type indicator, but this does not provide boundary information equivalent to that available to GMNO. Consequently, the main comparisons conflate architectural improvements with differences in input information. The shared-encoder ablation in Table 3 partially addresses this concern, but comparisons with equivalent physical inputs would better establish GMNO’s architectural advantages. Furthermore, all experiments concern Maxwell problems, while the mesh encoding and heterogeneous-input design appear applicable to other PDEs. Evaluation on non-electromagnetic mesh-based tasks, such as AirfRANS [3] or the Heat2d and NS2d-c settings in GNOT [4], would help substantiate this broader applicability. Otherwise, the claims should be qualified to reflect the current evidence from electromagnetic simulations.

[1] M. Lienen and S. Günnemann. Learning the Dynamics of Physical Systems from Sparse Observations with Finite Element Networks. ICLR, 2022.

[2] B. Zeng et al. PhyMPGN: Physics-encoded Message Passing Graph Network for Spatiotemporal PDE Systems. ICLR, 2025.

[3] F. Bonnet et al. AirfRANS: High Fidelity Computational Fluid Dynamics Dataset for Approximating Reynolds-Averaged Navier-Stokes Solutions. NeurIPS, 2022.

[4] Z. Hao et al. GNOT: A General Neural Operator Transformer for Operator Learning. ICML, 2023.

Questions for authors
  1. Could the authors report the zero-shot performance of the pretrained GMNO on the four held-out datasets before any fine-tuning? This would help distinguish its direct generalization capability from its ability to adapt using a small amount of target-task data.

  2. Fine-tuning provides only marginal gains on the seen tasks in Table 1. Could the authors investigate whether this reflects saturation after joint pretraining or substantial similarity among these tasks?

8 comments

We greatly appreciate the reviewer for the detailed and insightful feedback. We are grateful that the reviewer regards MaxwellBench as a valuable benchmark that covers multiple industrially relevant electromagnetic scenarios while retaining mesh structure and physical information. This is exactly what we built it for. We also thank the reviewer for recognizing GMNO's promising performance with cross-dataset pretraining and its few-shot adaptation to held-out tasks. Please find below our response to each point.

1. Novelty relative to FEN and PhyMPGN

Thank you for these references; we will cite and discuss both. We agree that both use structure beyond point coordinates, and we do not claim to be the first to use mesh cells or to encode boundary conditions. The difference lies in which mesh entities carry information, and which physical information they carry:

  • FEN triangulates the observation points and makes use of the dual graph of this mesh. A cell-wise MLP uses the vertex features, the cell centre and the local vertex coordinates to give out the embedding of each cell. This work then makes use of GNN framework, with its message-passing carried out through precomputed FEM basis-function integrals to process cell-wise information. The cell-wise information is then summed back to the vertices, so that the states and predictions live on the vertices. This work is tested on multiple spatio-temporal problems, with an explicit advantage on problems with unknown dynamics.
  • PhyMPGN also triangulates its nodes. It uses triangle geometry to build a cotangent-weighted discrete Laplace–Beltrami operator, which assumes a known diffusion term, and adds a learned correction, again acting on nodal states. It handles Dirichlet, Neumann, Robin and periodic conditions by padding ghost nodes in physical space (and in latent space for some types), together with a node-type flag. This acts on nodes of the outer boundary.
  • GMNO, by contrast, makes the 2D cells and the 1D segments the tokens themselves. Each token carries its own measure (area or length) and an embedding built by message passing over the solver's mesh connectivity. The prediction is made per cell, where the FEM field is defined.

GMNO has two token types at different levels:

  • 2D cell tokens carry materials (including a nonlinear B-H curve that changes from sample to sample), conductivity, complex sources and area, each with its own encoder.
  • 1D segment tokens carry the boundary-condition type, the normal, the materials on both sides and the length. They exist both on the domain boundary and on every internal material interface.

The two token types interact through cross-attention. Internal interfaces make up 66–98% of the 1D entities in MaxwellBench, and neither FEN nor PhyMPGN represents them or the materials on either side. In FEN's CylinderFlow experiment, boundary values are fixed outside the model. Both methods also learn the dynamics of one system per dataset, with coefficients fixed within the dataset and source terms learned implicitly rather than given as inputs. We will revise the related-work section and the contribution statement accordingly, crediting FEN for cell-based messages and PhyMPGN for explicit boundary-condition handling.

2. Fairness of the baseline comparison

All models receive the exact same simulation data. GMNO is given no extra information, and everything beyond the raw dataset comes from our algorithmic processing. The difference lies entirely in whether the model simplifies this information or processes it at full fidelity. In practical industrial scenarios, target fields are predicted on elements, while multi-level geometric and physical information is distributed across mesh nodes (points), 1D boundaries/interfaces (lines), and 2D elements (surfaces). Conventional point-cloud methods typically collapse each element, along with all its multi-level information, into a single point, coarsely compressing and flattening quantities from nodes, 1D lines, and 2D surfaces into uniform per-point channels. Under this simplification, explicit 1D edge attributes (normals, boundary types, and two-sided materials) cannot be directly merged into 2D element points without physical contradiction and ambiguity. GMNO instead distinguishes these entities and explicitly highlights boundary and interface transitions by constructing dedicated 1D and 2D tokens before the backbone. This shift from coarse simplification to fine-grained processing is the core contribution we want to emphasize.

To further illustrate the advantages of GMNO, part of our ablation results can be compared directly with the baselines: the vanilla baselines (which do not fully exploit the mesh information), the same baselines with full use of the mesh and physical information, and GMNO. Relative L2L_2 error (×10−2\times 10^{-2}):

BackboneInput encodingTrans UUInduc EI GapInduc LgEM C Wire
Transolver (Table 1)point cloud8.05.92.05.1
Transolver (Table 3)GMNO encoders6.14.21.02.4
GNOT (Table 1)point cloud12.68.83.77.5
GNOT, zero-padded linear attention (Table 3)GMNO encoders5.64.01.12.6
GMNOGMNO encoders3.83.81.02.3
  • The encoders. With GMNO's fine-grained encoders instead of coarse point-cloud compression, both baselines improve substantially: Transolver's error drops by 24–53% and GNOT's by 55–70%.
  • The backbone. With identical encoders, both baselines remain worse than GMNO, most clearly on Trans UU, the dataset with the most complex interfaces.

Thus, from the same underlying simulation information, the gains in Table 1 stem from both the fine-grained encoders and the backbone. Standard point-cloud baselines lose accuracy not because raw information is withheld, but because their overly simplified input compression blurs boundary and interface structures that matter for industrial devices.

3. Applicability beyond electromagnetics

We agree that our evidence is limited to 2D electromagnetics. Following your suggestion, we will qualify the claims in the paper accordingly and state the scope in the Limitations. We chose electromagnetics because it combines several properties that most existing benchmarks simplify away:

  • many material regions with nonlinear, sample-dependent constitutive curves;
  • many internal interfaces;
  • sources confined to sub-regions;
  • fields spanning several orders of magnitude.

AirfRANS and GNOT's Heat2d and NS2d-c are valuable and well-established benchmarks, and they already handle irregular geometry. However, they involve one or few materials with linear constitutive laws and few internal interfaces, and each dataset is generated by sampling within a single parameterized design space. They therefore lack the cross-scenario diversity (topologies, materials, and spatial scales) under the same governing PDEs that MaxwellBench provides to evaluate foundation model pretraining and transfer. Because of these simplifications, such classical benchmarks and baselines can hardly reflect the problems that neural operators face in real industrial settings. We proposed MaxwellBench and GMNO precisely to highlight this gap and thereby advance the practical deployment of neural operators. In future work, we will extend the evaluation to more application scenarios.

4. Zero-shot performance on held-out families

Thank you for this question, it separates two claims that the paper should keep apart. Relative L2L_2 error (×10−2\times 10^{-2}) on 960 test samples per held-out dataset:

ModelTrans LInduc EIInduc Lg GapEM C Chunk
GMNO 0.54B, zero-shot74.219.762.462.5
GMNO 0.1B, zero-shot100.620.362.861.0
GMNO 10.4M, trained from scratch on 240 samples33.120.129.810.2
GMNO 0.1B, fine-tuned on 240 samples4.73.47.33.6
GMNO 0.54B, fine-tuned on 240 samples3.53.05.93.3

Zero-shot transfer does not work for pretrained model, and we will state this clearly in the revision:

  • Three families introduce a new core template (Trans L), a new source geometry (EM C Chunk) or a new combination of core size and gaps (Induc Lg Gap). Specifically, Trans L has the most new elements: a new L-shaped core template, a new coil layout, and no core gap. EM C Chunk keeps the C-core of EM C Wire but replaces the array of round wires with one solid rectangular coil and allows a larger gap. Induc Lg Gap combines the large circular core of Induc Lg with the multiple gaps of Induc Sm Gap, at a different overall size. Even when parameters overlap, such changes of core or source topology alter both the field distribution and the mesh considerably. For example, EM C Chunk meshes have 1k–4k elements with a smallest element of 0.34 mm20.34\ \mathrm{mm}^2, against 5k–61k elements and 2.7×10−3 mm22.7\times 10^{-3}\ \mathrm{mm}^2 for EM C Wire. On these three families, all models reach 60–100% error without target data.
  • Only Induc EI, which recombines elements of pretraining families, has a moderate zero-shot error.
  • With 240 target samples, pretrained GMNO is about an order of magnitude below training from scratch.

Our claim is therefore data-efficient transfer, not zero-shot generalization. We will revise the wording.

5. Marginal gain of fine-tuning on seen tasks

In-distribution fine-tuning uses exactly the same 9,600 training samples per dataset that the model already saw during joint pretraining, with 10 epochs and a small learning rate. The small gain therefore indicates that joint pretraining works well: the model has already learned the features of the in-distribution datasets thoroughly and is not confused by being trained on many very different datasets at once. This also offers a useful reference for similar pretraining efforts, as it shows that joint pretraining alone is sufficient on the seen tasks.


We thank the reviewer again. The comments on input parity, on prior mesh-based methods and on zero-shot evaluation have made the paper's claims clearer and better supported.

Reviewer_1XhkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  1. MaxwellBench includes material B-H curves, interface conditions, mesh connectivity, and nonuniform element sizes. These detailed physical inputs make the dataset useful beyond the proposed model.

  2. The experiments are extensive. Parameter-matched baselines, three-seed results, architecture ablations, and a similarly sized pretrained Transolver provide useful support for the method.

Critical strengths
  1. MaxwellBench includes material B-H curves, interface conditions, mesh connectivity, and nonuniform element sizes. These detailed physical inputs make the dataset useful beyond the proposed model.

  2. The experiments are extensive. Parameter-matched baselines, three-seed results, architecture ablations, and a similarly sized pretrained Transolver provide useful support for the method.

Critical weaknesses
  1. The baselines do not receive the same physical information. Appendix C.1 states that Transolver, LNO, and GAOT receive only 2D inputs, while GMNO also receives 1D boundary and interface information. GNOT receives both, but uses padded attention. The common-embedding ablation is helpful, but it remains unclear how much of Table 1's gain comes from the architecture rather than these input differences.

  2. The generalization claim is broader than the evaluation supports. Errors are reported on log-normalized magnetic magnitudes, without errors in physical units or downstream engineering quantities. Two held-out families are subsets or extensions of existing design spaces. Please clarify these limits. The related work should discuss UPT and AB-UPT on scaling mesh-based simulation models, and PFEM on physics-driven pretraining with geometry, materials, and boundary conditions. These overlap with GMNO's mesh representation and physical inputs, so describing prior scaling work as mainly pixel-based is too broad.

Questions for authors
  1. Can the point-cloud baselines receive equivalent boundary and interface information? Are any existing comparisons available with matched inputs?
  2. Are errors in physical units or engineering quantities available? Please specify the log transform and how rotations and source flips are applied before taking magnitudes.
  3. Which changes in each held-out dataset are outside the pretraining design spaces, rather than subsets or extensions?
  4. How do the mesh representation and pretraining setup differ from UPT, AB-UPT, and PFEM?
4 comments

We sincerely thank the reviewer for the careful and constructive review. We are grateful that the reviewer finds MaxwellBench useful. Building it was a substantial part of this work: industrial electromagnetic simulation data with full mesh and material information are rarely publicly available, so we built MaxwellBench as a resource for the community, not only as a testbed for GMNO, and we hope it will be useful beyond our own model. We also thank the reviewer for recognizing the breadth of our experiments, including the parameter-matched baselines, three-seed results and architecture ablations. Below we answer the four questions in order, split over four comments.

1. Can the point-cloud baselines receive equivalent boundary and interface information?

All models receive the exact same simulation data. GMNO is given no extra information, and everything beyond the raw dataset comes from our algorithmic processing. The difference lies entirely in whether the model simplifies this information or processes it at full fidelity. In practical industrial scenarios, target fields are predicted on elements, while multi-level geometric and physical information is distributed across mesh nodes (points), 1D boundaries/interfaces (lines), and 2D elements (surfaces). Conventional point-cloud methods typically collapse each element, along with all its multi-level information, into a single point, coarsely compressing and flattening quantities from nodes, 1D lines, and 2D surfaces into uniform per-point channels. Under this simplification, explicit 1D edge attributes (normals, boundary types, and two-sided materials) cannot be directly merged into 2D element points without physical contradiction and ambiguity. GMNO instead distinguishes these entities and explicitly highlights boundary and interface transitions by constructing dedicated 1D and 2D tokens before the backbone. This shift from coarse simplification to fine-grained processing is the core contribution we want to emphasize.

To further illustrate the advantages of GMNO, part of our ablation results can be compared directly with the baselines: the vanilla baselines (which do not fully exploit the mesh information), the same baselines with full use of the mesh and physical information, and GMNO. Relative L2L_2 error (×10−2\times 10^{-2}):

BackboneInput encodingTrans UUInduc EI GapInduc LgEM C Wire
Transolver (Table 1)point cloud8.05.92.05.1
Transolver (Table 3)GMNO encoders6.14.21.02.4
GNOT (Table 1)point cloud12.68.83.77.5
GNOT, zero-padded linear attention (Table 3)GMNO encoders5.64.01.12.6
GMNOGMNO encoders3.83.81.02.3
  • The encoders. With GMNO's fine-grained encoders instead of coarse point-cloud compression, both baselines improve substantially: Transolver's error drops by 24–53% and GNOT's by 55–70%.
  • The backbone. With identical encoders, both baselines remain worse than GMNO, most clearly on Trans UU, the dataset with the most complex interfaces.

Thus, from the same underlying simulation information, the gains in Table 1 stem from both the fine-grained encoders and the backbone. Standard point-cloud baselines lose accuracy not because raw information is withheld, but because their overly simplified input compression blurs boundary and interface structures that matter for industrial devices.

2. Log targets, physical units and engineering quantities

Why a log target. There are two reasons.

  • Small fields matter for engineering, and a linear-scale loss barely sees them. The field spans several orders of magnitude within a single sample. In most datasets, the vast majority of the linear-scale ∥B∥2\lVert B \rVert^2 comes from the iron core, yet some physical quantities like inductance are very sensitive to much smaller ∥B∥2\lVert B \rVert^2 values in air gaps. In many datasets, about half of all target values are below 1% of the sample's maximum. Predicting every one of them as zero, i.e. with 100% error, would raise a linear-scale relative L2L_2 error by only 0.4–2.6%, which is about the size of our model's full-field ∣B∣\lvert B \rvert error. A linear-scale loss therefore hardly distinguishes a model that resolves the regions that determine inductance and winding loss from one that ignores them. On a log scale, a given relative error costs the same everywhere, so small fields must be predicted as accurately as large ones.
  • The log target is much easier to learn. On a linear scale the target is strongly right-skewed, so a few elements near the core dominate the loss and its gradients. After the transform, the target is nearly symmetric and lies in a fixed range of about [−3,4][-3, 4] for every dataset, which suits a standard regression head and a single model across datasets whose field levels differ.

The clamp affects only far-field air: elements below 5×10−4 T5\times 10^{-4}\ \mathrm{T} hold at most 0.2% of the stored energy and essentially none of the winding loss.

Evaluation in physical units. To check that this choice does not hide errors, we inverted the prediction and compared it with the raw, unclamped FEM solution on 960 test samples per dataset. We evaluated four engineering quantities:

  • Inductance: it can be calculated with energy method: W=12LI2W = \frac{1}{2}L I^2. The stored magnetic energy can be obtained with: W=∑eVe w(Be)W = \sum_e V_e\ w(B_e). Here VeV_e is the volume of each element, w(Be)=∥Be∥2/(2μ0)w(B_e) = \lVert B_e \rVert^2/(2\mu_0) for air and copper elements, and w(Be)=∫H dBw(B_e) = \int H\ \mathrm{d}B for ferromagnetic material elements in the core with their own nonlinear B-H curves.
  • Core loss: Steinmetz form ∑coreVe ∥Beβ∥\sum_{\mathrm{core}} V_e\ \lVert B_e^{\beta} \rVert with β=2.5\beta = 2.5, for ferrite cores.
  • Winding loss: Litz proximity-loss form ∑copperVe ∥Be∥2\sum_{\mathrm{copper}} V_e\ \lVert B_e\rVert^2.
  • Peak core flux density.

All results use the 0.54B model. The four in-distribution (ID) datasets are evaluated after pretraining: Trans UU (frequency-domain, Cartesian), Trans PQ (frequency-domain, axisymmetric), Induc Lg (no air gap) and EM C Wire (static C-core). The three held-out families are evaluated after fine-tuning on 240 samples: Trans L (new core template), Induc EI, and Induc Lg Gap (the hardest family).

Mean per-sample relative field error (%):

Datasetlog L2L_2 (paper metric)∣B∣\lvert B \rvert, element-weighted∣B∣\lvert B \rvert, volume-weighted
Trans UU (ID)2.22.23.2
Trans PQ (ID)2.02.24.9
Induc Lg (ID)0.81.11.2
EM C Wire (ID)1.41.41.6
Trans L (OOD)3.53.05.0
Induc EI (OOD)3.02.42.4
Induc Lg Gap (OOD)5.94.24.1

Mean per-sample relative error of engineering quantities (%):

DatasetInductance: meanCore lossWinding lossPeak core BB
Trans UU (ID)2.02.30.91.3
Trans PQ (ID)3.73.00.81.2
Induc Lg (ID)2.01.11.00.9
EM C Wire (ID)1.2––0.8
Trans L (OOD)2.53.41.91.8
Induc EI (OOD)1.92.01.41.7
Induc Lg Gap (OOD)5.56.65.63.0

As the two tables show, the errors in physical units and on the downstream engineering quantities that matter in practice are of the same size as the log metric reported in the paper, which indicates that our metric reflects the real accuracy well. Integrated quantities show little systematic bias, and small fields are predicted accurately as well.

Why magnitude targets suffice and how rotations/flips act before magnitudes and log normalization

Although magnetic flux density B\mathbf{B} is an axial vector field, downstream industrial quantities (inductance, core loss, winding loss, and peak flux density) depend only on the component magnitudes, ignoring vector directions and phase angles. Predicting scalar component magnitudes is therefore sufficient for downstream engineering evaluations. Crucially, every raw sample in MaxwellBench stores the full signed/complex vector components. During training, geometric and physical transformations are applied directly to the raw vector fields before extracting component magnitudes and applying the clamped log transform ∣∣B∣∣=log⁡(max⁡(∣∣B∣∣,5×10−4 T))/1.2+3.2||B|| = \log(\max(||B|| , 5\times 10^{-4}\ \mathrm{T})) / 1.2 + 3.2. This strict order of operations guarantees physical consistency without introducing unphysical artifacts.

These pre-magnitude transformations are used only in 0.54B pretraining with probability 0.4 per training sample:

  1. Translation. All node coordinates are shifted by a random offset of up to ±12 mm\pm 12\ \mathrm{mm}. For Cartesian samples, both directions are shifted; for axisymmetric samples, only the axial coordinate zz is shifted, so that rr keeps its physical meaning.
  2. Rotation (Cartesian only). The mesh coordinates, boundary normals, and raw signed/complex field vectors are rotated before magnitudes are taken. A 90∘90^\circ rotation maps (Bx,By)↦(−By,Bx)(B_x, B_y) \mapsto (-B_y, B_x) and swaps the resulting magnitude channels (∣Bx∣,∣By∣)(\lvert B_x\rvert, \lvert B_y\rvert).
  3. Source flip. The current density JJ and the raw vector field B\mathbf{B} are both negated before taking magnitudes. Because nonlinear B-H curves are odd, negating JJ physically negates B\mathbf{B} while leaving the magnitude target unchanged, teaching the model invariance to source polarity.

3. What lies outside the pretraining design space in each held-out dataset?

We agree that two of the four families recombine or extend pretraining families, and we will say so explicitly in the revision:

  • Trans L has the most new elements: a new L-shaped core template, a new coil layout, and no core gap.
  • EM C Chunk keeps the C-core of EM C Wire but replaces the array of round wires with one solid rectangular coil and allows a larger gap.
  • Induc Lg Gap combines the large circular core of Induc Lg with the multiple gaps of Induc Sm Gap, at a different overall size.
  • Induc EI has no distributed gaps and adds winding layers to the EI family.

Even when parameters overlap, a change of core or source topology changes both the field distribution and the mesh considerably. For example, EM C Chunk meshes have 1k–4k elements with a smallest element of 0.34 mm20.34\ \mathrm{mm}^2, against 5k–61k elements and 2.7×10−3 mm22.7\times 10^{-3}\ \mathrm{mm}^2 for EM C Wire. In industrial practice, air gaps likewise have a strong effect on both the mesh and the entire field distribution.

4. Relation to UPT, AB-UPT and PFEM

We will add all three, and we agree that describing prior scaling work as mainly pixel-based was too broad; we will remove that statement.

  • UPT encodes the simulation as a point cloud, aggregates it onto supernodes with radius-graph message passing, and further applies perceiver to encode into a fixed set of latent tokens. It explicitly does not use the original mesh connectivity, and it trains one model per dataset.
  • AB-UPT scales this approach to automotive CFD. Its input is coordinates only, and it uses separate geometry, surface and volume branches. Its surface tokens are close in spirit to our separate boundary tokens. However, they are positions used to predict surface quantities: they carry no boundary-condition type, normal or adjacent-material information, and the model has no internal material interfaces.
  • PFEM pretrains a Transolver on point clouds of FEM nodes, with input fields and boundary conditions stacked together as multiple channels. It is pretrained only with an FEM-discretized potential-energy loss, uses one small model per problem, and uses its prediction to warm-start a FEM solver.

GMNO differs from all three in four ways. It keeps 2D cells and 1D boundary/interface segments as separate token types, embedded through the message-passing according to mesh connectivity. Its 1D tokens encode boundary-condition type, length and the materials on both sides, including internal material interfaces, which make up 66–98% of the 1D entities in MaxwellBench. The inputs of 2D tokens include per-sample nonlinear B-H curves, material distributions and complex sources, each with its own encoder. Finally, it is pretrained with solution data jointly across ten device families and then transferred to held-out families. We will also cite AB-UPT in the Limitations as the main reference for extending GMNO to 3D.


We thank the reviewer again. These questions led to the physical-unit evaluation, the matched-input rows and a more accurate statement of the generalization claim, all of which we will include in the revision.

Reviewer_X9VTOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors presented a new method which encodes geometric and material information not only as node features but also as element features and used it inside a message passing algorithm to predict the physical response in a neural operator framework. The numerical examples on 2D problems show better performance compared to other methods.

Critical strengths

The dataset can help advance the research in a new domain.

It is shown that the proposed approach beats other methods with not an additional size or computational burden, and that it can be used as a retrained model effectively.

The idea of using element features is interesting because that information is already available from the mesh generation step.

Critical weaknesses

There is an additional cost associated with the element feature processing. It can limit the application when future mesh-free techniques offer better accuracies. Especially because they will bypass to need for a clean mesh.

Other methods have already successfully tested on 3D problems, and it is yet to be seen how this proposed approach work or how this would work on elasticity or navier stokes problems.

Questions for authors

N/A

1 comment

We sincerely thank the reviewer for the positive assessment and the insightful feedback. We are grateful that the reviewer sees MaxwellBench as a resource that "can help advance the research in a new domain". Building it was a substantial part of this work, because industrial electromagnetic simulation data with full mesh and material information are rarely publicly available, and we hope it will be useful to the community beyond our own model. We also thank the reviewer for finding the idea of using element features interesting. The reviewer's remark that this information "is already available from the mesh generation step" summarizes the motivation of GMNO well. Industrial FEM workflows already produce a mesh with element connectivity, element areas, material regions and boundary/interface labels, but most neural operators discard this structure when they convert the mesh into a point cloud. MaxwellBench was built to keep this structure, and GMNO is designed to use it. We are also glad that the reviewer found the pretraining results convincing. Below we respond to each of the weaknesses raised.

Computational cost of element-feature processing

We agree. Reading the element features costs essentially nothing, because connectivity, areas and edge lengths are direct outputs of the mesher and the FEM solver. The extra cost lies inside the model: GMNO keeps every 2D element and every 1D boundary/interface segment as a token, and it builds the embeddings by message passing over the mesh. This additional cost is what brings the accuracy gains in Table 1, so we regard it as a trade-off.

Dependence on a mesh and mesh-free alternatives

The reviewer's point is well taken. If future mesh-free approaches reach higher accuracy without needing a mesh, they would be a strong alternative to mesh-based pipelines built on FEM, and we see the two directions as complementary. GMNO targets the setting in which training data come from FEM, which is standard practice in electromagnetic design and, more broadly, in most industrial simulation workflows. In this setting, the mesh that produced each training sample is always available, and a new design is meshed anyway to obtain reference solutions. Mesh-free settings, however, would require further adaptation and remain to be studied. We will add the dependence on a mesh to the Limitations and list a mesh-free variant as future work, for example one in which a k-nearest-neighbour graph replaces mesh connectivity and boundary entities are extracted from the geometry.

Extension to 3D problems and other PDEs

We agree that this remains to be demonstrated, and we will restrict the claims of the paper to 2D electromagnetics. The representation itself is not specific to 2D or to Maxwell's equations:

  • 3D. The cells become 3D tetrahedra and the 2D segments become triangular faces, which carry the same attributes: boundary-condition type, normal, and the materials on both sides. There are also 1D edges to encode additional information.
  • Other PDEs. In elasticity or flow problems, the per-region inputs would be stiffness or viscosity instead of B-H curves.

The main obstacle is computational: industrial 3D meshes have millions of elements, as already noted in our Limitations. Works such as Transolver-3 and AB-UPT shows how large 3D meshes can be handled efficiently by reducing the number of tokens. The combination of such token-reduction techniques with our entity-based representation is the direction we plan to pursue next. We will add this discussion and these references to the Limitations section.


Thank you again for the supportive and constructive review. Your comments helped us state the cost and scope of GMNO more accurately.

Comments on the paper

1 comment

We thank the AC and all reviewers for their time. Below we briefly summarize the strengths, the main concerns, our responses and the reviewers' final feedback.

Strengths noted by reviewers

  • MaxwellBench is a valuable, industrially relevant benchmark that retains mesh structure and physical information, useful beyond our own model (X9VT, 1Xhk, ZkQ4).
  • Using element-level information is interesting and comes for free from mesh generation (X9VT).
  • Extensive experiments: parameter-matched baselines, three seeds, architecture ablations, a pretrained Transolver of similar size (1Xhk).
  • Strong performance with cross-dataset pretraining and few-shot transfer (X9VT, ZkQ4).

Main concerns and our responses

  • Novelty relative to prior mesh and entity-based methods (1Xhk, ZkQ4). FEN, PhyMPGN, UPT, AB-UPT and PFEM act on nodes or points. GMNO makes 2D cells and 1D segments the tokens, with modality-specific encoders (nonlinear B-H curves, sources, boundary types, two-sided materials at internal interfaces) and global heterogeneous attention. The contribution is this system together with the benchmark, the heterogeneous training paradigm and the evidence of scaling and transfer. We will cite all these works and state the novelty.
  • Fairness of the baseline comparison (1Xhk, ZkQ4). All models receive the same raw data, and the difference is how it is encoded. With GMNO's encoders, Transolver and GNOT improve by 24–70% but remain worse than GMNO, so the gains come from both encoders and backbone.
  • External validation beyond MaxwellBench (ZkQ4). On GNOT's Inductor2d, with the released data and GNOT's metric, GMNO reaches 1.15 / 1.64 / 0.84% (BxB_x / ByB_y / AzA_z), versus 25.99 / 38.81 / 1.65% for GNOT's unmodified official code. The released data store each interface node twice with two different values, which no point-wise model can represent. GNOT's final reported numbers are reproduced only after averaging the two values, and GMNO is better even against that easier (but wrong) setting.
  • Evaluation metric and practical relevance (1Xhk). Field errors in physical units and engineering quantities (inductance, core loss, winding loss, peak flux) on 7 ID and OOD datasets are of the same size as the paper's log metric (mostly 1–5%).
  • Scope of the generalization claims (1Xhk, ZkQ4). We added zero-shot results, clarified what is new in each held-out family and explained the small in-distribution fine-tuning gains as saturation after joint pretraining. We will describe the claim as data-efficient transfer, not zero-shot generalization.
  • Mesh dependence, cost, 3D and other PDEs (X9VT, ZkQ4). Element features come directly from the mesher, and FEM meshes are always available in industrial workflows. The representation extends to 3D and to other PDEs. We will restrict the claims to 2D electromagnetics and add mesh dependence and 3D scaling to the Limitations.

Reviewers' feedback

  • ZkQ4 stated that the Inductor2d evaluation and analysis substantially address the concerns and raised the rating from Weak Reject to Weak Accept, encouraging us to state GMNO's model-level novelty more clearly, which we will do.
  • X9VT (Clear Accept) and 1Xhk (Weak Accept) keep their positive ratings.

We believe all concerns have been addressed with new evidence, and we thank the AC for considering the full discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.