acceptodds
Under review as a conference paper at ICLR 2027

PHASE: Multi-Regime Modeling of Incompressible Magnetohydrodynamics

Abstract

Magnetohydrodynamics (MHD) is central to plasma modeling in astrophysics, space science, fusion, and engineering, but resolving multiscale MHD dynamics is computationally expensive. Machine-learning surrogates enable fast inference by learning reusable solution operators, yet existing models require separate training for each physical regime, limiting generalization across varying parameter settings. We introduce **PHASE**, a *PHysics-Adaptive Scalable operator with residual Error correction*, designed to model incompressible MHD across varying physical parameters with a single model. PHASE combines transfer learning, regime-aware adaptation, physics-centered learning, and residual refinement to improve both physical fidelity and generalization across MHD regimes. Together, these improvements achieve state-of-the-art prediction accuracy on two-dimensional MHD turbulence by reducing relative () errors on physical fields by more than an order of magnitude compared to prior MHD neural-operator baselines. Moreover, PHASE generalizes successfully to unseen parameter values without retraining, demonstrating the cross-regime adaptability expected from operator learning. We evaluate PHASE beyond pointwise prediction errors using derived physical fields, spectral analysis, and distribution statistics, consistently observing improved physical fidelity. We further show that our framework can accurately simulate MHD instabilities by testing it on the Kelvin-Helmholtz instability, demonstrating the robustness of our method.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operators, Physics, Physics Informed Neural Operators, Physics-informed neural networks, Magnetohydrodynamics, PDE, turbulence

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_sgm9OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Paper proposes PHASE, a neural operator for 2D incompressible magnetohydrodynamics that is a parameter-conditioned MHD operator.

Critical strengths
  1. Improvements in the prediction of magnetic-field and current-density quantities.

  2. Generalizes across multiple Reynolds-number regimes due to parameter conditioning.

Critical weaknesses
  1. Generalization to unseen parameter regimes is evaluated at only one point, Re=800, which was used during training.

  2. Limited to 2D flows, weak magnetic fields, and relatively low Re=Rm.

  3. Multiple physics components were combined, obscuring their relative benefits.

  4. Baselines are not directly comparable in terms of the training set, pretraining procedure, and computational budget. The baseline comparison is limited and does not include recent advances in neural operators, such as Transolver, Latent Mamba Operator, etc. Consider citing these recent works and including them as baselines with parameter conditioning.

Questions for authors
  1. Both the adapter up-projection and the gate’s final layer are initialized to zero, which would block gradient flow through the gated update. Is this intentional??

  2. How would a parameter-conditioned variant of DINO perform on the same multi-regime training distribution? Can you extend a similar setup to baselines based on Transolver and Latent Mamba Operator??

  3. Would the method generalize to other unseen Reynolds numbers, especially those outside the training distribution? If not, consider adding limitations.

  4. Can you evaluate the impact of each physics component independently (e.g., direct B prediction, Helmholtz projection, and derivative losses)?

  5. What is the computational overhead of the diffusion step, and does it contribute to the spectral power discrepancy? Also, can you compare it against [3]? A comparison of computational overhead with Transolver and Latent Mamba Operator would be useful to see whether we really need a diffusion backbone.

  6. Limited Novelty.

Reference:

  1. Wu, Haixu, et al. "Transolver: A fast transformer solver for pdes on general geometries." arXiv preprint arXiv:2402.02366 (2024).

  2. Tiwari, Karn, Niladri Dutta, and N. M. Krishnan. "Latent mamba operator for partial differential equations." arXiv preprint arXiv:2505.19105 (2025).

  3. Lippe, Phillip, et al. "Pde-refiner: Achieving accurate long rollouts with neural pde solvers." Advances in Neural Information Processing Systems 36 (2023): 67398-67433.

2 comments
AuthorsOpenReview ↗

Response to reviewer sgm9

We thank the reviewer for recognizing PHASE's gains in PDE prediction and cross-regime modeling. We address their concerns below.

1. Generalization and physical scope

We first clarify that Re=800Re=800 was not used in training. The MR training set is {80,200,400,650,1000,1500,2050,2750,3600,4500}\{80,200,400,650,1000,1500,2050,2750,3600,4500\}. We now additionally test generalization at unseen Re=1200Re=1200 and out-of-range Re=60,6000Re=60,6000, with no fine-tuning.

ReReDINO PrimaryMR PHASE PrimaryDINO DerivedMR PHASE Derived
600.9600.1214.4000.283
8000.2870.0420.9050.098
12000.2940.0550.7080.115
60000.5120.1270.8550.220

These results are shown in Fig. 5, Tab. 4, and App. A.7.

We agree that 3-D remains a limitation. The main study follows prior 2-D incompressible MHD work with Pm=1Pm=1. To broaden the scope, we added a strong-field, non-unity-PmPm case: Re=400, Rm=4000, Pm=10, Emag/Ekin=0.5Re=400,\ Rm=4000,\ Pm=10,\ E_{\rm mag}/E_{\rm kin}=0.5. PHASE reaches aggregate primary/derived errors of 0.20/0.380.20/0.38 and accurately captures the magnetically dominated evolution (App. A.9; Figs. 20$-$22).

2. Isolating the physics components

Fig. 2 and App. A.3 now isolate the individual physics additions:

Added componentPrimary field improvementDerived field improvement
Direct (u,B)(u,B) prediction19%52%
Helmholtz projection13%10%
Derived-field loss9%19%

The full ablation also isolates transfer learning, naive conditioning, gated adapters, and diffusion.

3. Baselines and comparability

As explained in our general response above, our goal is not a generic neural-operator benchmark but an AI-for-Science study for MHD. We therefore use tFNO-MHD (Rosofsky & Huerta, 2023) and DINO (Kacmaz, Huerta & Haas, 2025), the two directly comparable prior methods for the same 2-D incompressible MHD equations and DNS setting.

All models use DNS trajectories from the same Dedalus configuration, resolution, and temporal sampling. App. A.3 now reports splits, initialization/pretraining, and model sizes.

ModelDeterministicDiffusionTotal
tFNO16.97M--16.97M
DINO16.97M428.62M445.59M
MR scOT22.29M--22.29M
MR PHASE22.29M428.63M450.92M

Thus the gains are not explained by scale alone: DINO and PHASE have nearly identical total size, and their deterministic backbones are also comparable. Sec. 2 now additionally discusses Transolver, Latent Mamba Operator, PDE-Refiner, and recent MHD surrogates; parameter-conditioned MHD versions of these are valuable future work, but constitute a separate architecture-benchmarking study.

4. Adapter initialization

The implementation is correct; the original appendix didn't explicitly state the identity offset. The actual update is

hℓ+=hℓ+αℓgℓ(Φ)⊙Eℓ(hℓ)h_\ell^+=h_\ell+\alpha_\ell g_\ell(\Phi)\odot E_\ell(h_\ell),

where gℓ(Φ)=1+MLP(Φ)g_\ell(\Phi)=1+MLP(\Phi) and αℓ=1\alpha_\ell=1. Zero initialization of the gate MLP gives gℓ=1g_\ell=1, not zero, so the adapter up-projection receives nonzero gradients immediately; after it becomes nonzero, gradients reach the gate and down-projection. App. A.6.2 now state this explicitly.

The referred code in the text already contained the low level details: ReResidualAdapter, lines 70--141.

5. Diffusion cost and spectral tradeoff

Diffusion increases PHASE training cost to about 1.8$\times$ that of scOT alone and inference cost by about 50$\times$, while reducing aggregate prediction error by roughly 33%.

At Re=Rm=1000Re=Rm=1000:

ModelLow-kk uuLow-kk BBLow-kk ω\omegaLow-kk JJHigh-kk uuHigh-kk BBHigh-kk ω\omegaHigh-kk JJ
DINO0.0180.1760.0180.1784.5734.1544.5004.155
MR scOT0.0080.0160.0090.0170.1210.0490.6600.398
MR PHASE0.0050.0050.0050.0054.6863.0254.7993.051

Diffusion improves pointwise error and low-kk large/intermediate-scale accuracy, where MR PHASE is best across all fields. However, MR scOT is much better in the high-kk tail. Residual diffusion recovers unresolved structures while injecting excess small-scale spectral power. We now state this limitation explicitly in App. A.7.2 and Tab. 5.

6. Novelty

PHASE's novelty is the MHD-specific integration of cross-physics transfer, parameter-conditioned multi-regime adaptation, direct magnetic-field prediction, hard physical constraints, derived-field supervision, and residual correction. This combination substantially advances prior 2-D incompressible MHD surrogates and extends evaluation from decaying turbulence to instability-driven dynamics. See also our the general reply.

Thank you for the rebuttal. I will maintain my score.

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

My expertise is outside magnetohydrodynamics, so my assessment primarily focuses on the modeling and prediction of dynamical systems. PHASE adapts a pretrained fluid operator to two-dimensional incompressible MHD using parameter-conditioned adapters, direct velocity and magnetic-field prediction, Helmholtz projection, derivative-field supervision, and residual diffusion correction. A single model is trained across ten Reynolds numbers at fixed magnetic Prandtl number. Evaluation covers decaying turbulence and Kelvin–Helmholtz instability, including field errors, spectra, distribution statistics, and instability evolution.

Critical strengths

The improvements extend beyond primary-field accuracy. At Re = 1000, the reported current-density error decreases from 1.561 for DINO to 0.108 for PHASE, with substantially better current-density kurtosis and magnetic-energy evolution. These diagnostics make the physical benefit more convincing than an aggregate error comparison alone.

The paper provides a useful engineering study of transferring a fluid foundation model to MHD. The staged comparisons support the value of pretraining and parameter conditioning, while the KH experiments demonstrate applicability to a distinct evolution problem.

Critical weaknesses

The adapter initialization appears inconsistent with its intended training. Appendix A.6.2 states that both the adapter up-projection and the gate network's final layer are zero-initialized. Under the stated update, h←h+g(Φ)⊙E(h)h \leftarrow h+g(\Phi)\odot E(h), if both branch outputs are zero, each branch receives zero gradient through their product. An omitted nonzero gate activation, offset, or different initialization could resolve this, but the current description does not explain how the central adaptation mechanism learns.

The evidence for cross-regime adaptation is narrower than the framing suggests. The only reported unseen Reynolds number is 800, inside the training range; all experiments keep Re = Rm, and Appendix A.2 describes an extremely weak magnetic field. Moreover, the multi-regime model receives data from ten regimes and additional fine-tuning, so its advantage over single-regime training does not by itself isolate the benefit of parameter adaptation. The naive-conditioning comparison is useful, but the data and optimization controls need to be explicit.

Questions for authors
  1. What are the exact gate activation, biases, and initialization used in the experiments? Please reconcile them with the simultaneous zero initialization in Appendix A.6.2 and explain how the adapter branches receive nonzero gradients.

  2. Were naive conditioning and gated adapters compared using identical trajectories, initialization checkpoints, training steps, and model-selection rules? Please also specify the trajectory-level training, validation, and test splits.

  3. Are results available at additional held-out Reynolds numbers within the stated training range, using frozen weights? A small sweep would clarify whether the reported generalization extends beyond the single Re = 800 interpolation case.

1 comment
AuthorsOpenReview ↗

Response to reviewer KYtk

We thank the reviewer for their careful reading and for recognizing that the improvements extend beyond aggregate field error to current density, magnetic-energy evolution, and instability dynamics. We address the technical concerns below.

1. Adapter initialization and gradient flow

We agree that the original description in App. A.6.2 was incomplete. The implementation used in all experiments includes an identity offset in the gate:

$h_\ell^{+} =h_\ell+ \alpha_\ell g_\ell(\Phi)\odot E_\ell(h_\ell),$

where αℓ=1\alpha_\ell=1 and the actual gate is gℓ(Φ)=1+MLP(Φ).g_\ell(\Phi)=1+MLP(\Phi).

The gate MLP and adapter up-projection are zero-initialized, but this gives gℓ=1g_\ell=1, not 0. Therefore the adapter up-projection receives nonzero gradients immediately. On the first step, the gate branch itself has zero gradient because Eℓ(hℓ)=0E_\ell(h_\ell)=0; once the up-projection moves away from zero, gradients flow into the down-projection and gate. Thus the initialization produces a staged warm start rather than a deadlock. The revised App. A.6.2 now states this explicitly.

The exact implementation is:

conditioning = self._conditioning_features(re, rem)
gate = 1.0 + self.gate_net(conditioning).to(hidden_states.dtype)

update = self.up(
    self.act(
        self.down(
            self.norm(hidden_states)
        )
    )
)

return hidden_states + self.adapter_scale * gate * update

The submitted anonymous repository already contained this implementation: ReResidualAdapter, lines 70--141.

The experimental settings are also now explicit: bottleneck dimension 6464, GELU activations, biases enabled, adapter scale =1=1, zero-initialized adapter up-projection and final gate layer, and standard initialization elsewhere.

2. Controlled comparison: naive vs. gated conditioning

Yes. The naive-conditioning and gated-adapter models were compared under the same experimental controls:

  • identical trajectory sets;
  • identical 800/100/100 train/validation/test split per regime;
  • same initialization checkpoint;
  • same losses and normalization;
  • same balanced batching and optimizer setup;
  • ∼100\sim 100 training epochs;
  • model selection by minimum normalized validation loss;
  • final reporting only on the held-out test set.

These details are now stated in App. A.3.

The controlled ablation at Re=1000Re=1000 shows that naive multi-regime conditioning improves the transferred model, but gated adaptation provides a substantially larger gain:

ModelPrimary field rel. L2L_2Derived field rel. L2L_2
+ transfer learning0.1420.782
+ naive MR conditioning0.1190.647
+ gated adapters0.0670.364

This isolates the benefit of the parameter-conditioned adapter design beyond simply adding multi-regime data.

3. Generalization to additional unseen Reynolds numbers

We clarify that Re=800Re=800 was not used in training. The training set is {80,200,400,650,1000,1500,2050,2750,3600,4500}.

We have now expanded the frozen-weight generalization study to:

  • unseen interpolation: Re=800,1200Re=800,1200;
  • unseen extrapolation: Re=60,6000Re=60,6000.

No fine-tuning was performed for these cases.

Re=RmRe=RmDINO PrimaryMR PHASE PrimaryDINO DerivedMR PHASE Derived
600.9600.1214.4000.283
8000.2870.0420.9050.098
12000.2940.0550.7080.115
60000.5120.1270.8550.220

These results are shown in Fig. 5 and Tab. 4, with corresponding field, spectral, and PDF analyses in App. A.7.

4. Beyond the original weak-field Pm=1Pm=1 setting

The main experiments intentionally matched the prior MHD literature with Re=RmRe=Rm (i.e. Pm=1Pm=1). To test whether PHASE remains effective outside this setting, we added a strongly magnetized case with Re=400,Rm=4000,Pm=10,Re=400,\quad Rm=4000,\quad Pm=10, and initial Emag/Ekin=0.5E_{\rm mag}/E_{\rm kin}=0.5. PHASE achieves aggregate primary- and derived-field relative L2L_2 errors of 0.20 and 0.38, respectively, while accurately reproducing the transition to a magnetically dominated state. Spectra, PDFs, time evolution, and field structures are presented in App. A.9, Figs. 20–22.

We hope these revisions resolve the implementation ambiguity and strengthen the evidence that the multi-regime gains come from controlled parameter adaptation rather than merely additional data or optimization.

Reviewer_UY71OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors address the problem of modeling multi-scale magnetohydordynamics The authors introduce PHASE a model developed using transfer learning, regime-aware adaptation, physics-centered learning, residual refinement. The model outperforms reported baselines

Critical strengths
  • this paper investigates the challenge of modelling multiscale magnetohydrodynamics, advancements here are of clear benefit for several domains
Critical weaknesses
  • I found the paper difficult to read. I think this is primarily because of the organization. For example, the method section describes training procedures and interpretations of the empirical results in some places.
  • The contributions weren’t clear contributions to me, for example, contribution (1) is that transfer learning was used. (4.1) expands on this describing the training procedure and incorporation of a conditional framework. I imagine these contribute to improved performance; however, they don’t seem to be particularly novel contributions.
  • I’m not sure the contributions are validated by the empirical results. For example the transfer learning procedure is not contrasted with other transfer learning procedures.
  • I’m not deeply familiar with the domain, however, I suspect the related work (only two methods are cited for ML surrogates for MHD) is missing important citations
Questions for authors
  • It would be good to see a comparison of model sizes.
1 comment
AuthorsOpenReview ↗

Response to reviewer UY71

We thank the reviewer for recognizing the importance of multiscale MHD modeling and its relevance across several scientific domains. We address each concern below.

1. Clarifying the novelty of the contributions

We agree that transfer learning by itself is not the contribution. The novelty is the adaptation of modern ML components to a setting where they have not previously been demonstrated together: multi-regime incompressible MHD. (See the authors’ general reply above.)

In particular, PHASE:

  • transfers pretrained fluid-dynamics representations from POSEIDON into a coupled velocity$-$magnetic field system;
  • learns magnetic-field evolution and velocity--magnetic coupling that are absent from the fluid pretraining task;
  • introduces Re/RmRe/Rm-conditioned adapters so that one shared model operates across multiple physical regimes;
  • combines direct magnetic-field prediction, hard divergence-free projection constraint, derived-field supervision, and residual refinement.

Thus, the contribution is not a new generic transfer-learning algorithm, but a cross-physics, physics-adapted MHD framework. We will revise the introduction and Sec. 4.1 to make this distinction explicit.

2. Are the contributions empirically validated?

Yes. Our goal is not to compare different transfer-learning algorithms, but to test whether fluid pretraining is useful for MHD and whether each additional PHASE component provides a measurable benefit. The revised Fig. 2 and App. A.3 now provide a more granular ablation.

Model stagePrimary field rel. L2L_2Derived field rel. L2L_2
scOT from scratch0.4681.770
+ POSEIDON transfer0.1420.782
+ naive multi-regime conditioning0.1190.647
+ gated adapters0.0670.364
+ direct u,Bu,B prediction0.0540.176
+ Helmholtz projection0.0470.159
+ derived-field loss0.0430.129
+ diffusion (PHASE)0.0280.085

This directly isolates the value of transfer learning, conditioning, physics-centered changes, and residual refinement. Importantly, the scOT backbone alone underperforms DINO; the gains arise from the complete PHASE framework rather than simply choosing a transformer architecture.

3. Related work

We acknowledge the original related-work discussion was too narrow. It emphasized tFNO and DINO because they are the two directly comparable prior methods for the same 2-D incompressible MHD equations and DNS setting. We have now expanded Sec. 2 to include less directly related work, including Kim et al. (2024), Duarte et al. (2025), Chang et al. (2025), Kacmaz & Huerta (2026), and Pothula & Kumar (2026), as well as general neural operators such as Transolver, Latent Mamba Operator, and PDE-Refiner.

4. Model size and computational cost

We agree and now report parameter counts and computational costs explicitly in App. A.3.

ModelDeterministicDiffusionTotal
tFNO16.97M--16.97M
DINO16.97M428.62M445.59M
MR scOT22.29M--22.29M
MR PHASE22.29M428.63M450.92M

The performance gain is therefore not explained by parameter count alone. The deterministic tFNO and MR scOT backbones are of comparable scale (17M vs. 22M), yet the physics-adapted scOT is dramatically more accurate after transfer, conditioning, and MHD-specific improvements. Likewise, DINO and PHASE have nearly identical total size once diffusion is included (∼\sim 446M vs. ∼\sim 451 M), while PHASE achieves much lower errors.

We hope these revisions clarify both the AI-for-Science scope and the empirical support for each component of PHASE.

5. Organization and readability

Given the space constraint, we focused on the MHD background and the key ideas behind the empirical performance improvements, connecting them to the ML-specific and physics-specific changes. Low-level details were deferred to the Appendix and code repository. We have reorganized the main paper and will try to use more general-audience language in the final version. Training schedules, initialization details, optimizer settings, and implementation choices have been moved to App. A.4$-A.6,whileempiricalinterpretationisconcentratedinSec.5andApp.A.7A.6, while empirical interpretation is concentrated in Sec. 5 and App. A.7-$A.9. This separation is intended to make the scientific logic of the framework easier to follow.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General reply to all reviewers

We thank the reviewers for their feedback. Before addressing the reviewer-specific comments below, we clarify the intended scope and evaluation lens of this work, as several concerns appear to arise from viewing PHASE primarily as a pure ML contribution. PHASE is instead an AI-for-Science framework for MHD that combines modern ML advances with MHD-specific physical structure. While transfer learning, adapters, physics-informed losses, and diffusion are not individually new, integrating and adapting them to multi-regime MHD is novel in this setting and enables the gains reported here. All revisions addressing reviewers' comments are highlighted in blue in the revised manuscript.

Scope and significance

MHD is central to astrophysics, fusion, space physics, and engineering, yet high-fidelity simulations are extremely expensive. Recent large-scale MHD simulations can require $ \sim 8 \times 10^7 $ CPU core-hours for a single parameter setting, making repeated exploration across initial conditions and physical parameters costly.

ML surrogates can amortize this cost through reusable solution operators, but prior MHD surrogates are generally trained separately for each physical regime. Our central question is therefore whether modern ML components can be physics-adapted into a single MHD surrogate that remains accurate across regimes and generalizes to unseen parameters. We encourage the paper to be evaluated through this ML-for-MHD lens: whether it materially advances predictive accuracy, physical fidelity, and cross-regime usefulness in an emerging AI-for-Science domain.

Baselines and contribution

We compare against tFNO-MHD [1] and DINO [2] because they are the two directly comparable prior methods developed for the same 2-D incompressible MHD equations and DNS setting. DINO achieves the strongest previously reported results in this setting. General-purpose PDE architectures such as Transolver, Latent Mamba Operator, and PDE-Refiner are now discussed in Sec. 2, but have not been developed for this MHD problem; constructing parameter-conditioned MHD versions of each would constitute a separate architecture-benchmarking study.

PHASE contributes:

  • Cross-physics transfer: POSEIDON fluid representations provide a velocity prior while magnetic evolution and velocity--magnetic coupling are learned from MHD data.
  • Multi-regime adaptation: (Re/RmRe/Rm)-conditioned residual adapters allow one shared model to operate across multiple regimes instead of retraining one model per setting.
  • Physics-centered improvements: direct magnetic-field prediction, hard divergence-free projection constraint, and explicit supervision of vorticity and current density improve small-scale physical fidelity.
  • Residual refinement: diffusion predicts residual errors unresolved by the deterministic operator.

Together these yield more than an order-of-magnitude improvement in several field-error metrics compared with prior state-of-the-art MHD methods, while extending evaluation beyond pointwise accuracy to spectra, PDFs, derived fields, divergence preservation, and Kelvin--Helmholtz instability dynamics (Secs. 4--5; Figs. 2--7).

[1] Rosofsky & Huerta (2023), “Magnetohydrodynamics with Physics Informed Neural Operators.”
[2] Kacmaz, Huerta & Haas (2025), “Resolving turbulent magnetohydrodynamics: a hybrid operator-diffusion framework.”

Clarifications and revisions

Several concerns arose from factual misunderstandings:

  1. Unseen-ReRe generalization. Re=800Re=800 was not used in training. The training set is {80,200,400,650,1000,1500,2050,2750,3600,4500}. We additionally evaluate frozen-weight generalization at unseen Re=1200Re=1200 and out-of-range Re=60,6000Re=60,6000. PHASE remains substantially more accurate than DINO in all four unseen regimes without retraining (Fig. 5; App. A.7; Tab. 4).

  2. Adapter initialization. The implemented gate is gℓ(Φ)=1+MLP(Φ),g_\ell(\Phi)=1+MLP(\Phi), so zero initialization of the conditioning MLP gives gℓ=1g_\ell=1, not 0, and the adapter up-projection receives nonzero gradients immediately. We have made Eq. 7 and App. A.6.2 explicit; the submitted anonymous code already contained the correct implementation.

  3. Physical breadth. The main Pm=1Pm=1 setting follows prior work. We now add a strongly magnetized non-unity-PmPm case with Re=400,Rm=4000,Pm=10Re=400, Rm=4000, Pm=10, showing PHASE remains effective when viscous and magnetic dissipation scales differ (App. A.9).

  4. Ablations and controls. We expanded the ablations to isolate gains from transfer learning, naive vs. gated conditioning, direct (u,B)(u,B) prediction, Helmholtz projection, derived-field losses, and diffusion refinement (Fig. 2; App. A.3). We also report model sizes, training/inference costs, data splits, optimization controls, and the distinct low-kk and high-kk effects of diffusion (Tabs. 1$-$5).

Below, we address each reviewer comment and point to the corresponding revisions and new experiments in detail.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.