- 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
- MaxwellBench provides a valuable benchmark for neural operators by covering multiple industrially relevant electromagnetic scenarios while retaining mesh structure and physical information.
- GMNO achieves promising predictive performance through cross-dataset pretraining and demonstrates few-shot adaptation to related electromagnetic tasks.
Critical weaknesses
- 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.
- 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
-
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.
-
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
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 error ():
| Backbone | Input encoding | Trans UU | Induc EI Gap | Induc Lg | EM C Wire |
|---|---|---|---|---|---|
| Transolver (Table 1) | point cloud | 8.0 | 5.9 | 2.0 | 5.1 |
| Transolver (Table 3) | GMNO encoders | 6.1 | 4.2 | 1.0 | 2.4 |
| GNOT (Table 1) | point cloud | 12.6 | 8.8 | 3.7 | 7.5 |
| GNOT, zero-padded linear attention (Table 3) | GMNO encoders | 5.6 | 4.0 | 1.1 | 2.6 |
| GMNO | GMNO encoders | 3.8 | 3.8 | 1.0 | 2.3 |
- 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 error () on 960 test samples per held-out dataset:
| Model | Trans L | Induc EI | Induc Lg Gap | EM C Chunk |
|---|---|---|---|---|
| GMNO 0.54B, zero-shot | 74.2 | 19.7 | 62.4 | 62.5 |
| GMNO 0.1B, zero-shot | 100.6 | 20.3 | 62.8 | 61.0 |
| GMNO 10.4M, trained from scratch on 240 samples | 33.1 | 20.1 | 29.8 | 10.2 |
| GMNO 0.1B, fine-tuned on 240 samples | 4.7 | 3.4 | 7.3 | 3.6 |
| GMNO 0.54B, fine-tuned on 240 samples | 3.5 | 3.0 | 5.9 | 3.3 |
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 , against 5k–61k elements and 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.
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:
GMNO has two token types at different levels:
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.