acceptodds
Under review as a conference paper at ICLR 2027

Toward a Physics-Informed TokaMind: Learning Magnetohydrodynamic Equilibrium Laws

Abstract

Although foundation models for scientific domains promise scalable prediction across diverse modalities, they often lack explicit constraints to governing physical laws. In tokamak dynamics, this is critical since surrogates must predict magnetic equilibrium features essential for safe operation, but predictions from standard data-driven learning approaches are not guaranteed to respect the governing laws of plasma confinement (which is worsened by the scarcity of experimental fusion data compared to other domains). We present a methodology to augment the multi-modal foundation model TokaMind with physics-informed losses derived from the Grad-Shafranov equation, the cornerstone constraint on tokamak equilibrium. We introduce two loss formulations based on strong-form and weak-form residuals, and develop new evaluation metrics that additionally target physical consistency and operator-space accuracy. Fine-tuning experiments based on the TokaMark benchmark reveal that physics-informed regularization substantially reduces mutual consistency errors between predicted flux and current density, improves flux accuracy in operator space, and softens small-scale structural disruptions. These improvements come at a modest cost in signal-space normalized RMSE, indicating a fundamental trade-off between data fidelity and physical consistency. Our work connects foundation models and physical reasoning, providing both a reusable template for physics-informed training in scientific machine learning, and evidence that governing laws improve the reliability of equilibrium predictions for high-stakes plasma physics applications. Our open-sourced implementations will be made available upon acceptance.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Grad-Shrafranov equation, Physics-Informed Machine Learning, fusion plasma physics, foundation model, TokaMind

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

This study proposes a physics-informed learning framework for magnetohydrodynamic (MHD) equilibrium prediction in tokamak fusion devices using the multimodal Transformer architecture of TokaMind. In the experiments, the TokaMind architecture is trained from random initialization with an additional physics loss derived from the Grad–Shafranov (GS) equation. Two formulations of the physics loss are introduced: a strong-form loss that directly evaluates the differential-equation residual and a weak-form loss based on a discretized variational formulation. The training objective combines the data and physics losses with coefficients of 0.9 and 0.1, respectively. Experiments on FAIR-MAST data across six flux reconstruction and reconstructive forecasting tasks show that the GS operator-space error EψE_\psi decreases by approximately 50% across all six tasks, while the GS self-consistency error EselfE_{\mathrm{self}} decreases by approximately 50% for the four tasks that predict both poloidal magnetic flux and toroidal current density. In contrast, the flux prediction error NRMSEψ\mathrm{NRMSE}_\psi increases by approximately 4–18%. The methodological contributions of this study include the integration of GS-based physics regularization into the TokaMind architecture, the strong- and weak-form loss formulations, and the introduction of physics-aware evaluation metrics. The main empirical finding is that improvements in physical self-consistency and operator-space accuracy do not necessarily translate into improved pointwise prediction accuracy under the evaluated conditions.

Critical strengths

The main strength of this paper lies in its demonstration of the practical applicability of physics-informed learning to a realistic and complex scientific problem. Specifically, the authors incorporate physics constraints derived from the Grad–Shafranov equation into the Transformer-based TokaMind architecture and evaluate their effectiveness in magnetic equilibrium prediction using multivariate diagnostic data from the MAST tokamak experiment, made available through FAIR-MAST. Although incorporating known PDEs into neural network loss functions is a well-established approach, its effectiveness under realistic conditions involving incomplete experimental data, consistency across multiple physical quantities, and competition with data-fitting objectives is not necessarily guaranteed. Demonstrating that this established framework can successfully operate in such a complex, real-world scientific setting is an important contribution to the AI for Science community, and I believe that such practical demonstrations will become increasingly valuable in the future.

Critical weaknesses

The main weakness of this paper is that both PDE-residual-based regularization and weak-form variational formulations are well-established methodologies in Physics-Informed Machine Learning. The paper introduces no new learning principles, optimization methods, network architectures, or theoretical properties. Moreover, magnetic equilibrium reconstruction using actual tokamak diagnostic data with the Grad–Shafranov (GS) equation imposed as a physics constraint has already been investigated in previous studies, including Liao et al. (2026), which is cited in this paper. The main empirical finding, namely the trade-off between physical consistency and data prediction accuracy, is also a well-known issue associated with competing loss terms in PINNs and can hardly be considered a novel insight. Therefore, the novelty of this work is limited to extending existing methodologies to TokaMind, a multimodal Transformer, and multiple prediction tasks. I recognize that ICLR also values empirical studies that do not introduce new algorithms. However, it remains unclear what generalizable insights this study provides to the machine learning community beyond existing knowledge.

Questions for authors

Q1. The data-loss-only model achieves a lower NRMSE but a larger GS residual. What factors do you think are responsible for this result, such as errors in the EFIT reference solution, approximation errors arising from DCT and finite-difference discretization, or limitations of the quasi-static equilibrium assumption?

Q2. Are the following typographical errors or inconsistencies?

  1. §3.2.2, opening paragraph: The text states "GS diagnostics of Section 3.2.3 (namely EselfE_{\mathrm{self}}, EψE_\psi, and the roughness deviation)," but the roughness deviation is defined in §3.2.4. This should be "Sections 3.2.3 and 3.2.4."

  2. Appendix B.2, Eq. (32): The right-hand side of ∥GSres∥2=∑i(GSres(i))2\|GS_{\mathrm{res}}\|_2 = \sum_i (GS_{\mathrm{res}}(i))^2 represents the squared norm. Either ∥⋅∥22\|\cdot\|_2^2 should be used, or the square root should be taken on the right-hand side.

  3. Reproducibility Statement: The evaluation metrics are referenced as being defined in "Appendix B.2," but Appendix B.2 discusses the loss function metrics. The correct reference should be "Appendix E.2."

  4. §3.2.1: Although the text states that RMSEk,i\mathrm{RMSE}_{k,i}, RMSEk\mathrm{RMSE}_k, WkW_k, σk\sigma_k, and SS are being defined, the corresponding equations are missing, and these quantities are not explicitly defined. The missing equations should be restored.

  5. Abstract and §1: The Abstract refers to "Fine-tuning experiments," and §1 refers to "tuning experiments," whereas §4.1 states that the models are trained from random initialization.

  6. §4.2: The text states "lowest NRMSEψ\mathrm{NRMSE}_\psi in most tasks," but Table 2 shows that the data-only model achieves the lowest NRMSEψ\mathrm{NRMSE}_\psi in all six tasks.

  7. §4.2: The text states "NRMSEJϕ\mathrm{NRMSE}_{J_\phi} is essentially unchanged throughout," but Task 2-4 shows an improvement of 11.7%, which is also inconsistent with the description in the Conclusion.

  8. §4.2, Limitations: "”collapsed”" should be "“collapsed”".

5 comments
AuthorsOpenReview ↗

Responses to Reviewer WBCS

We thank the reviewer for valuing practical demonstrations of physics-informed learning on real experimental data, and for the very careful reading that found the listed inconsistencies (all of which are now corrected). Below we address raised weaknesses and questions, with more details provided in the general response.

AuthorsOpenReview ↗

Response to Weakness 1

Weakness 1 - Novelty and generalizable insight. We agree that PDE-residual and weak-form regularization are established, and that GS-constrained reconstruction from real diagnostics has been studied, e.g., by Liao et al. (2026). Our revised paper now presents itself as an empirical study and positions itself against Liao et al. more explicitly. In Liao et al., the authors train a bespoke, single-device PINN in which the GS equation is the modelling target. In our work, GS terms are auxiliary objectives in a multi-modal, multi-task model with partial output specifications and a compressed output basis. We think the setting changes the conclusion in a way that matters to the ML community. The trade-off we observe is not the generic competition between loss terms seen in PINNs, where an exact solution would satisfy both terms but optimization conflicts prevent reaching it. In our case, the targets themselves satisfy the physics, and the compromise comes from approximating in a fixed low-dimensional output space under two different norms (general response, G2). The general lesson is that for any model that predicts in a compressed basis (e.g., DCT, PCA, autoencoder latent), the output representation is part of the design of a physics-informed loss. A differential operator changes which errors the compression leaves, so the codec can create a data/physics trade-off that the physics alone would not. We will state this lesson explicitly in the introduction and conclusion, together with the observation (in Appendix E.1) that higher codec fidelity also passes more measurement-noise content to the operator.

AuthorsOpenReview ↗

Response to Question 1

Question 1 - Why does the data-only model have lower NRMSE but larger GS residual? Thank you for this question. Of the factors you list, we believe the main one is the approximation error from the output representation, i.e., the DCT truncation, not the finite-difference discretization. The decoded ψ\psi lies in the span of KgK_g between 3939 and 4848 DCT3D coefficients. The data loss picks the element of that span closest to ψtrue\psi^{true} w.r.t. the selected loss metric. Truncation errors in that element are small in amplitude but concentrated at small spatial scales, and Δ∗\Delta^* amplifies exactly these, so EψE_\psi is large. The GS term picks a different element, closer w.r.t. the loss metric in operator space, at the cost of a slightly higher NRMSEψNRMSE_\psi (44 to 1818\\%, Table 2) (G2). The other factors play smaller roles. EFIT solves GS with its own discretization and parametrized source profiles, so the reference pair satisfies our discrete operators only up to discretization error. This sets a non-zero floor, which we will report (Appendix E.2 already uses it as a check). The quasi-static assumption is built into EFIT (shared by targets and losses), and it does not explain the gap between our models. Measurement noise affects the targets only indirectly through the reconstruction, and its high-frequency part is mostly removed by the codec at threshold 0.99 (Appendix E.1). Consistent with this, the flux-error increase is concentrated outside the JϕtrueJ^{true}_{\phi}-based reference limiter (G7). We will add this discussion to Section 4.2.

Reviewer_9MGNOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates physics-informed learning for tokamak equilibrium reconstruction by incorporating Grad–Shafranov (GS) equation constraints into the TokaMind architecture. The authors introduce strong-form and weak-form physics-based loss formulations, together with additional evaluation metrics for physical consistency and operator-space accuracy. Experiments on six TokaMark tasks show substantial reductions in GS consistency and operator-space errors, while generally increasing signal-space prediction errors. The study highlights a trade-off between data fidelity and physics-based regularization.

Critical strengths

Well-motivated application of physical constraints. The paper addresses an important limitation of data-driven plasma equilibrium reconstruction: accurate numerical predictions do not necessarily satisfy governing physical laws. Incorporating GS constraints is meaningful for improving physical consistency.

Critical weaknesses

a. Limited methodological novelty. The proposed approach primarily integrates established physics-informed regularization techniques into an existing TokaMind architecture. Both strong-form and weak-form PDE constraints are well studied, and the paper does not introduce a substantially new learning mechanism. Furthermore, the experiments train the architecture from scratch rather than demonstrating physics-informed adaptation of a pretrained foundation model, limiting the support for its broader foundation-model motivation.

b. Insufficient evidence of improvements beyond the optimized physics constraints. Although the proposed method substantially reduces GS-related residual metrics, these metrics are closely related to the physical constraints used during training. Meanwhile, signal-space prediction accuracy consistently deteriorates, and the exploratory downstream transfer results do not demonstrate a general advantage. Consequently, it remains unclear whether the improved physical consistency translates into more accurate or reliable equilibrium predictions in practice.

Questions for authors
  1. Can the authors provide more independent evidence that the improved GS consistency translates into physically more accurate equilibrium reconstruction or better downstream performance, beyond metrics closely related to the training constraints?

  2. Given that the experiments train TokaMind from random initialization, how would the proposed physics-informed losses perform when applied to a pretrained TokaMind foundation model? Does the approach improve transferability or preserve previously learned representations?

  3. How sensitive are the reported trade-offs to the physics-loss weight? Could repeated-seed experiments and a systematic weight analysis establish whether similar physical improvements can be achieved with less degradation in signal-space accuracy?

5 comments
AuthorsOpenReview ↗

Responses to Reviewer 9MGN

We thank the reviewer for the provided feedback and for recognizing why physical constraints matter for data-driven equilibrium reconstruction. Below we address raised weaknesses and questions, with more details provided in the general response.

AuthorsOpenReview ↗

Response to Weakness A

Weakness A - Methodological novelty and foundation-model motivation. We agree that strong- and weak-form residual regularization are established techniques, and the revised manuscript says so explicitly. We have rewritten the contributions to present the paper as an empirical study of these techniques in a setting where their behavior is not obvious: a multi-modal, multi-task model with (i) partial output specifications, (ii) sparse native-space supervision, and (iii) outputs decoded from a compressed DCT basis. The method-level contributions are the residual taxonomy of null/no-anchor/L-anchor/R-anchor GS residuals that makes the objective reduce correctly across tasks with and without JϕJ_\phi outputs, and an evaluation protocol whose GS diagnostics use a discretization that neither training operator uses. The main empirical insight is the explanation of the data/physics compromise as a consequence of approximating in a compressed output space under two different loss metrics (general response, G2). Concerning the foundation-model motivation, the reviewer is right that Section 4.1 trains from scratch. The manuscript revision removes "fine-tuning" mentions and presents the work as a study of the training objective for the TokaMind architecture, with adaptation of the pretrained model as future work (G1).

AuthorsOpenReview ↗

Response to Weakness B and Question 1

Weakness B / Question 1 - Evidence beyond the optimized constraints. We agree that the reported GS diagnostics are closely related to the training terms, and the manuscript revision says so directly in Sections 3.2 and 4.2, not only in the Limitations. We will no longer claim that our results show more accurate or more reliable equilibria, and the revised claim is limited to physical consistency and operator-space accuracy. We add a clarification of what EψE_\psi represents physically: the error in the current density implied by the predicted flux via Ampère's law (G3). We also note that the diagnostics are averaged over the full eroded limiter, which is larger than the plasma region on which the GS loss acts for the (ψ,Jϕ)\left(\psi, J_\phi\right) tasks (G6). A regional breakdown (G7) shows that the flux-error cost lies mostly outside the JϕtrueJ^{true}_{\phi}-based reference limiter, where no physics term acts. We also state clearly that the downstream-transfer study is exploratory and shows no general advantage. Evaluation of quantities derived from ψpred\psi^{pred} (poloidal field, boundary geometry) is listed as the main open question.

Reviewer_c7vgOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper adds physics-informed losses to the existing TokaMind architecture for predicting tokamak magnetic flux and current density. The losses encourage agreement with reference flux derivatives and consistency between predicted flux and current. The authors compare strong- and weak-form versions on reconstruction and forecasting tasks using experimental data with equilibrium-reconstructed targets.

Critical strengths

I like the idea of a whole paper studying one regularization term. I think that this can be very worthwhile, if one goes into enough detail about it. I thought the improvements in current forecasting were interesting.

Critical weaknesses

(1) I feel like the main problem is that you have this new metric, you're training against something very close to that metric, and then you show that metric going down. I would like to see whether it helps with something that isn't directly related to that consistency. I appreciate that they use held-out data and a different discretization, but my sense is that this is still testing essentially the same relationship.

Now, the authors do show that forecasting the current improves in certain settings, which I thought was interesting. The downside of this is that the flux error gets worse, and in the companion configuration, those current improvements mostly disappear. So what I would like to see is whether the magnetic field or the plasma boundary calculated from the flux actually gets more accurate. But this separate downstream geometry test doesn't really tell us that. Also, as I understand it, making flux and current match is not enough to establish full force balance, which seems stronger than what this metric can tell us in Appendix E.2.

(2) The other thing I could not understand is the claim that there's a fundamental conflict between the data and the physics. Because, to my understanding from reading a bit of the cited literature, the target maps were reconstructed using the same equation they're using for regularization. So that information is actually there in the data. For flux-only tasks, they compare derivatives of the prediction with derivatives of the same reference flux. Reproducing that reference exactly would satisfy both losses. So, I don't see why the observed compromise establishes this as a fundamental conflict. It could come from, e.g., their compressed outputs, model capacity, or something else.

And when they predict both flux and current, they combine that derivative supervision with a penalty making the two predictions match. I would like to see the derivative supervision on its own, so we can tell whether making the two outputs match actually helps.

Questions for authors

(1) Could you provide the physical evaluations and derivative-only comparison discussed above? (2) The model is trained from scratch in Section 4.1, so I'm not sure why this is presented as adapting a foundation model. Can the authors clarify this?

5 comments
AuthorsOpenReview ↗

Responses to Reviewer c7vg

We thank the reviewer for a precise and technically informed review, and for recognizing the value of studying one regularization term in depth. Several of the reviewer's points are correct, and we have changed the manuscript accordingly. Below we address raised weaknesses and questions, with more details provided in the general response.

AuthorsOpenReview ↗

Response to Weakness 1

Weakness 1 - Evaluation close to the training objective. We agree with this. EselfE_{self} and EψE_\psi are consistency diagnostics closely tied to the no-anchor and L-anchor terms. Scoring every model with one finite-difference operator that no loss optimizes, on held-out shots and, for the (ψ,Jϕ)(\psi,J_\phi) tasks, over a region larger than the one on which the GS loss acts, reduces this circularity but does not remove it (G3, G6). We will therefore no longer present these reductions as evidence of improved physical accuracy. We do note that EψE_\psi is not merely a training residual: through Ampère's law, it is the error in the current density implied by the predicted flux, i.e., the current that a downstream code would see (G3).

We also agree with the three related points. First, the NRMSEJϕNRMSE_{J_\phi} gains in Tasks 2-4/2-6 come with a higher NRMSEψNRMSE_\psi and largely disappear in the companion configuration (Table E.2), so we will no longer present them as a consistent benefit. Second, Tasks 1-2/2-2 test whether learned representations transfer, not a boundary computed from ψpred\psi^{pred}; we will say so and list the ψpred\psi^{pred}-derived field and LCFS/X-point geometry as open items. Third, agreement between Δ∗ψ\Delta^*\psi and μ0RJϕ\mu_0 R J_\phi is Ampère's law, which is necessary but not sufficient for force balance; we will describe EselfE_{self} accordingly (G3).

AuthorsOpenReview ↗

Response to Weakness 2

Weakness 2 - "Fundamental conflict". We agree and withdraw the claim (full argument in G2). As the reviewer notes, for ψ\psi-only tasks both losses are minimized by ψpred=ψtrue\psi^{pred}=\psi^{true}. The compromise comes from what the model can output, close to the reviewer's own suggestion (compressed outputs, capacity): TokaMind decodes ψ\psi from only KgK_g between 39 and 48 DCT3D coefficients, so it cannot reproduce ψtrue\psi^{true} exactly, and the data loss and the GS loss in operator space, which penalizes small-scale errors much more heavily, then favor different approximations. Our tables support this: the data-only baselines of Tables 2 and E.2 differ only in the codec threshold, and the higher-fidelity codec lowers their NRMSEψNRMSE_\psi by 1818 to 4141\\% but changes their EψE_\psi by less than 55\\%, whereas the GS term halves EψE_\psi at a fixed codec.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response to all reviewers (Part 3/3)

G7. Where the flux-error cost arises (Table 1). To investigate the spatial distribution of prediction errors, we computed the flux MSE separately on the full grid and inside the limiter (Table 1). On the full grid, the physics terms increase MSEψMSE_\psi by 3−403-40\\%; inside the limiter, the change is only −7-7 to +12+12\\%, and with SGS the error is even lower than the data-only baseline in three of six tasks. MSEJϕMSE_{J_\phi} changes identically in both regions, as expected since Jϕ≈0J_\phi \approx 0 outside the plasma boundary. Thus, most of the cost in the flux map lies outside the limiter, near the poloidal-field coils, where no physics term acts and the plasma GS relation does not apply; because the DCT3D codec is global, enforcing GS consistency inside the limiter shifts the fit outside it. This also supports G2, showing that data fidelity is largely preserved within the physics-constrained region.

Table 1. Test results at explained-energy threshold of 0.990.99. MSE is evaluated over the full grid and within the non-eroded limiter using the same pooled aggregation: the sum of squared prediction errors divided by the number of valid scalar prediction–reference pairs in the selected region, pooled across all evaluated windows, time slices, and shots. Shots with more valid elements receive greater weight. MSEψMSE_{\psi} is reported in units of 10−5 (Wb/rad)210^{-5}\,(\mathrm{Wb}/\mathrm{rad})^2, and MSEJϕMSE_{J_{\phi}} in units of 109 (A/m2)210^{9}\,(\mathrm{A}/\mathrm{m}^{2})^2. [NOTE: Everything else in the original caption remains unchanged.]

Bold = best, italic = second best. ↑ / ↓ give the percentage change relative to the first loss (LD) in each task group.

Task, LossMSEψMSE_{\psi} full gridMSEψMSE_{\psi} limiterMSEJϕMSE_{J_{\phi}} full gridMSEJϕMSE_{J_{\phi}} limiterEselfE_{\mathrm{self}}EψE_{\psi}ϵrough\epsilon_{\mathrm{rough}}
1-3, LD18.594415.3065------0.3452+0.1911
1-3, LDSGS19.0795 (↑2.6%)15.1470 (↓1.0%)------0.1816+0.0295
1-3, LDWGS19.8131 (↑6.6%)15.5739 (↑1.7%)------0.1670+0.0164
1-4, LD2.67991.79731.24461.73730.33030.3310+0.1989
1-4, LDSGS3.4024 (↑27.0%)1.8450 (↑2.7%)1.2570 (↑1.0%)1.7547 (↑1.0%)0.1711 (↓48.2%)0.1746 (↓47.3%)+0.0677
1-4, LDWGS3.7531 (↑40.0%)2.0021 (↑11.4%)1.2807 (↑2.9%)1.7864 (↑2.8%)0.1624 (↓50.8%)0.1662 (↓49.8%)+0.0526
1-6, LD2.96362.09511.63022.29580.32910.3303+0.1971
1-6, LDSGS3.6894 (↑24.5%)2.1438 (↑2.3%)1.6403 (↑0.6%)2.3096 (↑0.6%)0.1702 (↓48.3%)0.1756 (↓46.8%)+0.0665
1-6, LDWGS4.0472 (↑36.6%)2.3478 (↑12.1%)1.6882 (↑3.6%)2.3792 (↑3.6%)0.1607 (↓51.7%)0.1667 (↓49.5%)+0.0494
2-3, LD14.205011.5061------0.3380+0.1806
2-3, LDSGS15.8131 (↑11.3%)12.2979 (↑6.9%)------0.1830 (↓45.9%)+0.0129
2-3, LDWGS15.8554 (↑11.6%)12.1745 (↑5.8%)------0.1644 (↓51.4%)-0.0040
2-4, LD4.46883.47131.22851.76820.33150.3319+0.1944
2-4, LDSGS4.9143 (↑10.0%)3.2262 (↓7.1%)0.9721 (↓20.9%)1.3970 (↓21.0%)0.1765 (↓46.8%)0.1793 (↓46.0%)+0.0620
2-4, LDWGS5.2464 (↑17.4%)3.3689 (↓2.9%)0.9885 (↓19.5%)1.4213 (↓19.6%)0.1667 (↓49.7%)0.1702 (↓48.7%)+0.0445
2-6, LD4.52853.53241.24121.78570.33470.3341+0.1924
2-6, LDSGS5.1365 (↑13.4%)3.4419 (↓2.6%)1.1365 (↓8.4%)1.6355 (↓8.4%)0.1750 (↓47.7%)0.1789 (↓46.5%)+0.0554
2-6, LDWGS5.5420 (↑22.4%)3.6837 (↑4.3%)1.1770 (↓5.2%)1.6943 (↓5.1%)0.1651 (↓50.7%)0.1698 (↓49.2%)+0.0388
AuthorsOpenReview ↗

General response to all reviewers (Part 2/3)

G3. Claims narrowed to physical consistency; what EψE_{\psi} measures. We agree that EselfE_{self} and EψE_{\psi} are closely related to the training terms, and we will say so in the main text, not only in the Limitations. Three clarifications will be added as follows:

  • Both diagnostics use a centred finite-difference operator ΔFD∗\Delta^*_{\mathrm{FD}} that neither loss optimizes, so a model is never scored with the operator it was trained on. As Reviewer c7vg notes, this reduces but does not remove the circularity. In addition, the diagnostics are averaged over the fixed, one-cell-eroded limiter, which is larger than the reference plasma region on which the GS loss acts (G6) for the four (ψ,Jϕ)\left(\psi, J_\phi\right) tasks.

  • EψE_{\psi} has a direct physical reading: by Ampère's law, ΔFD∗(ψtrue−ψpred)\Delta^*_{\mathrm{FD}}(\psi^{true}-\psi^{pred}) is, up to discretization error and the factor −μ0R-\mu_0 R, the error in the toroidal current density implied by the predicted flux map. This is the current density that any downstream use of ψpred\psi^{pred} (e.g., an equilibrium or stability code) would see. Neither point turns EψE_{\psi} into an independent accuracy metric, and we will not present it as one. In this context, we will also correct our description of EselfE_{self}: low EselfE_{self} certifies consistency of (ψ,Jϕ)(\psi, J_\phi) with Ampère's law, which is necessary but not sufficient for MHD force balance. Full force balance also requires JϕJ_\phi to have the flux-function structure Rp′(ψ)+FF′(ψ)/(μ0R)R p'(\psi) + F F'(\psi)/(\mu_0 R), which none of our metrics tests; the EFIT targets have this structure by construction, but our metrics do not test whether the predictions inherit it.

  • Finally, we will soften the claim that our results show "improve[d] reliability of equilibrium predictions" (abstract, l. 29--30) and list evaluation of ψpred\psi^{pred}-derived quantities (poloidal field, boundary/X-point geometry) as an open item.

G4. Positioning and contributions. We agree that strong- and weak-form residual losses are established techniques, and we will say so explicitly. What is different is how we use them: an anchored-residual formulation that lets a single GS loss act as an auxiliary objective across tasks with different output specifications (including tasks that do not predict JϕJ_\phi), in a multi-modal, multi-task model with a compressed output basis. Nonetheless, we will rewrite the contribution list so the paper reads as an empirical study with the following contributions: (a) a residual taxonomy (null / no-anchor / L-anchor / R-anchor) of GS residuals, adaptable to partial output specifications in a multi-task benchmark so that the same objective reduces correctly when JϕJ_\phi is not an output; (b) an evaluation protocol with GS diagnostics that use a discretization different from either training operator; (c) a controlled comparison of strong- and weak-form GS terms on real FAIR-MAST data with sparse supervision and a compressed output space; and (d) the projection-based account of the data/physics compromise (G2), which applies to any physics-informed loss acting on a model that makes predictions in a compressed basis.

G5. Statistical scope and presentation. All the reported results are single-seed with fixed weights. We will state this in Section 4.2 next to the results and remove language that implies robustness ("suggests the core findings are robust'', l. 469). We will fix every typographical and cross-reference issue raised by reviewer WBCS, restore the missing NRMSE definitions in Section 3.2.1, and make a consistency pass over notation (anchor names, operator symbols) to address the raised presentation issues.

G6. Correction to the evaluation-region description. We identified a mismatch between the evaluation region described in Section 3.2.2 and the implementation used for Tables 2 and E.2. The reported GS diagnostics and roughness values were computed over the fixed, one-cell-eroded limiter E\mathcal{E} for all tasks, without the additional reference-current mask stated in Eq. (18). We will correct the region definition and the corresponding text and captions. This is a correction to the description only, with all the reported numerical results remaining unchanged, and it also clarifies the relation to training: for tasks predicting both ψ\psi and JϕJ_\phi, this evaluation region differs from the reference JϕtrueJ^{true}_{\smash \phi}-based mask (Appendix B.1) used by the GS loss during training. With this, the post-training diagnostics also cover the region between the plasma boundary and the limiter in which no physics term is evaluated during training (where the GS loss reduces to 0 as a consequence of the cleaning mask MM).

AuthorsOpenReview ↗

General response to all reviewers (Part 1/3)

General response to all reviewers‾\underline{\text{General response to all reviewers}}

We thank the three reviewers for their careful and constructive reviews. We acknowledge that they have identified several places where some of our claims go beyond what our results support, and thus we agree to revise those points accordingly. Below we summarize the main ways to address raised concerns and give more details in the individual responses to each reviewer. We also report one correction that we found ourselves while revising the manuscript (namely, point G6).

G1. Scope and framing: training objective, not foundation-model adaptation. All experiments in Section 4.1 train the TokaMind architecture from random initialization. Therefore, "fine-tuning" in the abstract and "tuning" in Section 1 are wrong, and we will correct them. We trained from scratch with the purpose to isolate the effect of the physics term from pretraining-dependent confounders. The published pretrained TokaMind also uses a different output codec (explained-energy threshold 0.999, Appendix E.1), which would make the physics term act on a different output space. We will revise the abstract, introduction, and contributions so that the paper is presented as a study of Grad-Shafranov-informed training objectives for a multi-modal, multi-task TokaMind-style model, and will mention that a physics-informed adaptation of a pretrained checkpoint is envisioned as future work. The only experiments that use warm starts are the downstream-transfer study (Appendix E.4), and we will label them as such.

G2. The "fundamental trade-off" claim is withdrawn and replaced by a mechanism. Reviewers c7vg and WBCS are correct. The EFIT targets are themselves Grad-Shafranov (GS) solutions (up to discretization and within the plasma boundaries), so reproducing ψtrue\psi^{true} and JϕtrueJ_\phi^{true} exactly using TokaMind would minimize the data and physics losses at the same time (providing some special conditions are satisfied such as proper hyper-parameter tuning of TokaMind's codec, and a suitable cleaning mask for JϕJ_\phi is available).

The observed trade-off is therefore not a fundamental conflict between data and physics, and we will remove that wording (abstract, l. 26; conclusion, l. 459--463). We will instead give the following explanation, which (i) is specific to our setting, and (ii) we think is the more useful finding. TokaMind decodes each gridded output from a small set of retained DCT3D coefficients, namely, Kg=39K_g = 39 for ψ\psi in Tasks 1-x, and Kg=48K_g = 48 in Tasks 2-x (Table E.1). The decoded field therefore lies in a fixed low-dimensional linear subspace V=range(D)\mathcal{V} = \mathrm{range}(D), and individual targets ψtrue\psi^{true} generally lie outside it. On V\mathcal{V}, the data loss selects the best approximation in a (masked) loss metric. The L-anchored GS term selects the best approximation w.r.t. the loss metric in operator space R(M⊙T(⋅))\mathcal{R}\left( M \odot T(\cdot) \right) with T∈{Δ∗,W}T \in \{\Delta^*, W\} and R\mathcal{R} a suitable loss metric, which weights small-scale components much more heavily. Unless ψtrue∈V\psi^{true} \in \mathcal{V}, the two best approximations differ, and the convex combination in Eq. (26) moves the solution between them. A rise in NRMSEψNRMSE_{\psi} together with a fall in EψE_{\psi} is exactly what this predicts. Finite capacity, finite training, and the mismatch between the EFIT discretization and our operators add to the effect. In short, the compromise comes from approximating in a compressed output space under two different norms, not from the physics contradicting the data. We will state this as an interpretation consistent with our results, not as a proven decomposition, and we will discuss its implications: the trade-off should shrink as V\mathcal{V} grows and as model capacity improves, and the choice of output representation is part of the design of a physics-informed loss.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.