- 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 decreases by approximately 50% across all six tasks, while the GS self-consistency error decreases by approximately 50% for the four tasks that predict both poloidal magnetic flux and toroidal current density. In contrast, the flux prediction error 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?
-
§3.2.2, opening paragraph: The text states "GS diagnostics of Section 3.2.3 (namely , , 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."
-
Appendix B.2, Eq. (32): The right-hand side of represents the squared norm. Either should be used, or the square root should be taken on the right-hand side.
-
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."
-
§3.2.1: Although the text states that , , , , and are being defined, the corresponding equations are missing, and these quantities are not explicitly defined. The missing equations should be restored.
-
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.
-
§4.2: The text states "lowest in most tasks," but Table 2 shows that the data-only model achieves the lowest in all six tasks.
-
§4.2: The text states " is essentially unchanged throughout," but Task 2-4 shows an improvement of 11.7%, which is also inconsistent with the description in the Conclusion.
-
§4.2, Limitations: "”collapsed”" should be "“collapsed”".
5 comments
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.
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 lies in the span of between and DCT3D coefficients. The data loss picks the element of that span closest to w.r.t. the selected loss metric. Truncation errors in that element are small in amplitude but concentrated at small spatial scales, and amplifies exactly these, so 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 ( to , 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 -based reference limiter (G7). We will add this discussion to Section 4.2.
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.