{"id":"436592d0-bfd0-4e8a-ac65-79ff0bdd152e","arxiv_id":"2412.00087","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The Onion architecture, which multiplies backbone features with diagnostic response-matrix encodings and adds a line-integral loss, modestly reduces reconstruction error on synthetic and experimental fusion data.","lead":"Researchers propose a physics-informed deep learning architecture, Onion, that feeds diagnostic response matrices into neural networks and adds a line-integral consistency loss to reconstruct 2D plasma profiles from line-integral measurements such as soft X-ray cameras. It reports lower reconstruction errors on synthetic and experimental tokamak data, with millisecond inference times, which could support real-time plasma control.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PILF's claimed 'correction of inversion-algorithm errors' is not established: on experimental data E2 falls by construction while E1 rises several-fold, and no comparison against known true profiles separates genuine correction from merely enforcing x=R·y.","rationale":"The reader's verdict is CONDITIONAL and already identifies the forward-model exactness as the weakest assumption. My stress-test agrees with the overall conditional assessment but sharpens the concern: the decisive failure mode is not only mis-calibration of R, it is that the paper's headline claim about PILF 'reducing errors associated with inversion algorithms' is untested against any ground truth. On experimental data, the PILF-driven E2 decrease is tautological because E2 is essentially the same functional as loss2, while E1 degrades substantially. The paper's interpretation of the single MHD sample in Figure 19 is anecdotal and assumes the discrepancy between the target profile and the input is an inversion-algorithm error, rather than a consequence of an inexact forward model or of overfitting to loss2. The synthetic datasets are the natural place to settle this, but the paper does not run the controlled experiment with noisy labels that would mimic the experimental regime. Therefore the central claim should be endorsed only conditionally, pending that test. I do not raise a non-finding because the concern is real and concrete; nevertheless, the underlying architecture work is coherent and the comparative tables for PI and Softplus on synthetic data provide partial support, so the appropriate move is to require the missing validation rather than reject the paper outright.","tokens_in":20834,"tokens_out":2232,"duration_ms":24114,"concrete_test":"Use the synthetic EAST and HL-2A generators with known true profiles and inject zero-mean noise Delta_i into Eq. (2-1) so that epsilon_bar matches the experimental values (about 5e-2). Train VggOnion_PI and ResOnion_PI with loss1 only and with PILF using c1=0.618 and c1=1.0. Measure E1 against the known true profiles. If PILF does not reduce E1 relative to loss1-only, or if the profiles with lower E2 are farther from the true profiles in E1, then the claim that PILF corrects inversion-algorithm errors is unsupported. As a secondary check, perturb the response matrices R by a small calibration error and verify whether the PILF gains persist or reverse.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that the physics-informed loss function 'reduces the errors associated with inversion algorithms' requires that the forward model in Eq. (2-1), x_i = R_i·y + Delta_i, is accurate and that penalizing Eq. (3-5) moves predictions toward the true plasma profile. Neither condition is validated. Table 5 shows that adding PILF on Exp_EAST raises E1 from 0.50e-2 to 1.35e-2 (VggOnion_PI) and from 0.45e-2 to 1.37e-2 (ResOnion_PI), and on Exp_HL-2A from 0.21e-2 to 1.94e-2 and from 0.21e-2 to 1.93e-2, respectively. E2 drops because it directly penalizes the same back-projection mismatch that loss2 minimizes, so that improvement is by construction. The paper interprets the resulting disagreement with inversion-algorithm targets as correction of inversion errors, but that interpretation rests on the unverified assumption that the experimental targets are wrong in exactly the direction predicted by loss2. No experimental ground-truth profile exists in the study, and the synthetic datasets, where truth is known, are only used in the noise-free regime where loss1 alone suffices. The per-dataset tuning of c1, the absence of uncertainty information, and the lack of independent calibration of the response matrices R further weaken the inference. Thus the load-bearing gap is not the architecture itself but the unsupported causal claim that minimizing E2 corrects inversion errors rather than trading profile fidelity for consistency with a possibly miscalibrated forward model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a physics-informed deep learning architecture, 'Onion', for reconstructing 2D plasma emissivity profiles from line-integral soft X-ray diagnostic measurements. The architecture feeds a positional-encoded response-matrix 'physical information' (PI) side chain into a backbone network (VGG or ResNet) via element-wise multiplication, and augments the standard supervised loss with a line-integral consistency loss (loss2) that penalizes the difference between the back-projection of the network output and the measured chords. The authors evaluate the contributions of PI, a Softplus activation in the final fully connected layers, and the physics-informed loss function (PILF) on two synthetic datasets (noise-free, generated by their own forward model) and two experimental datasets from EAST and HL-2A, reporting reductions in the average relative reconstruction error E1 and the back-projection error E2. The paper also provides execution-time comparisons showing that the surrogate models are orders of magnitude faster than the F-B and NSGPT inversion codes.","tokens_in":21076,"tokens_out":3634,"duration_ms":32943,"significance":"If the claims were fully validated, the Onion architecture would be a useful, backbone-agnostic wrapper for embedding diagnostic response matrices into neural-network reconstructions, with potential impact on real-time plasma control and between-shot analysis. The paper contributes a synthetic data generator, two experimental datasets, and a clear description of the architecture and loss design, as well as a code/data availability commitment. The strongest asset is the architectural idea of multiplying learned features by fixed physical-information maps, which is simple and plausibly transferable. However, the current evidence is not sufficient to support the central claim that PILF corrects inversion-algorithm errors: the E2 improvement is largely a consequence of minimizing the same mismatch that E2 measures, and the paper lacks ground-truth validation, uncertainty quantification, multiple-seed statistics, and an independent calibration of the forward model. The significance is therefore conditional on additional validation.","major_comments":[{"comment":"The reported reduction in E2 after adding PILF is largely by construction, because loss2 in Eq. (3-5) penalizes the same back-projection mismatch that E2 in Eq. (4-2) measures. Consequently, the abstract's claim that PILF 'reduces the errors associated with inversion algorithms' is not supported by E2 alone. Table 5 shows that the E2 improvement is accompanied by a large degradation in E1, e.g., on Exp_HL-2A E1 rises from 0.21e-2 to 1.94e-2 for VggOnion_PI and from 0.21e-2 to 1.93e-2 for ResOnion_PI. Without a known ground truth (or an independent diagnostic) demonstrating that the reduced E2 corresponds to a more accurate plasma profile, the interpretation that PILF corrects inversion errors remains unsupported; please provide validation on noisy synthetic data where the true profile is known, or temper the claim to consistency with the forward model.","section":"Section 4.3, Eq. (3-5) and Eq. (4-2)"},{"comment":"The entire physical-consistency mechanism rests on the accuracy of the forward model x_i = R_i * y + Delta_i (Eq. 2-1). The paper provides no independent calibration or validation of the response matrices R_i against a known source, and the experimental datasets contain no uncertainty information. If the response matrices are miscalibrated or the systematic error Delta_i is not negligible, both the PI side chain and loss2 will actively pull predictions toward an incorrect measurement model. The authors should include a sensitivity analysis to perturbations in R (or a calibration check against an emissivity standard) and should state explicitly how the absence of Delta_i uncertainty affects the strength of the physical constraint.","section":"Section 2.1-2.2 and Section 3.2, Eq. (2-1)"},{"comment":"All reported metrics come from a single training run per configuration, with no error bars, no repeated-seed statistics, and no statistical significance tests. Several of the claimed improvements are very small in absolute terms, such as the average E1 reduction of 0.06e-2 on experimental datasets from adding PI (Table 3), which may be within run-to-run variation. Please report mean and standard deviation over at least five independent training runs, and where possible compare with the baseline inversion algorithms (F-B code and NSGPT) on the same held-out test samples rather than only through execution time and aggregate dataset errors.","section":"Section 4, Tables 3-5"},{"comment":"The hyperparameter c1 that weights loss2 is set per dataset and per model (0.618 for Exp_EAST, 1.0 for Exp_HL-2A) with no sensitivity analysis, no principled selection criterion, and no discussion of how sensitive the E1/E2 trade-off is to this value. Since c1 directly controls the balance between matching the inversion-algorithm target and matching the raw measurements, the paper should report results across a range of c1 values and justify the chosen values; otherwise the strong E2 reduction on Exp_HL-2A (from 5.38e-2 to 0.76e-2) could simply reflect an overly large c1 that sacrifices profile fidelity for measurement consistency.","section":"Section 4.3, Table 5"}],"minor_comments":[{"comment":"The theoretical motivation for multiplication over addition or concatenation relies on a single-neuron example and a statement that 'empirical testing of both approaches revealed suboptimal model performance'; please provide or cite the actual comparison results, otherwise the claim is not verifiable.","section":"Section 3.2, Eq. (3-1) to (3-3)"},{"comment":"The execution-time comparison is not controlled: F-B code timings are cited from a laptop with an i7-8750H, NSGPT from an A100 server, while the surrogate models are timed on a different desktop with an RTX 3050; the reported speedup factors therefore mix hardware and method differences.","section":"Table 2"},{"comment":"The training loss loss1 is a mean squared error while the evaluation metric E1 is a mean absolute relative error; the paper should state explicitly why the training objective and the evaluation metric differ, and whether an alternative loss (e.g., relative L1) was explored.","section":"Equation (3-4) vs. Eq. (4-1)"},{"comment":"The paper correctly acknowledges that the experimental datasets lack uncertainty information, but the statement that this omission 'precludes the conduct of uncertainty-related research' should be reconciled with the strong claims about improving reconstruction accuracy, since the same lack of uncertainty also limits the conclusions that can be drawn about experimental performance.","section":"Section 2.1"},{"comment":"The synthetic datasets consist of simple circular emissivity profiles, which are not representative of the peaked, hollow, or MHD-perturbed profiles found in real tokamak discharges; the transferability of the synthetic-data results to experimental conditions is therefore limited and should be stated as a caveat.","section":"Section 2.2 and Section 4.1"},{"comment":"The description of the PILF result on Exp_EAST says the in-out asymmetry 'seems to be in the opposite direction with respect to the target profile'; this qualitative statement should be either quantified (e.g., through a line-out comparison) or softened, since it is central to the claim that PILF corrects a core-region error.","section":"Figure 19 and Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the architectural idea is interesting, but the central claim about PILF correcting inversion errors is not yet established. The most important missing pieces are: (i) validation against known ground truth on noisy synthetic data, (ii) repeated-seed statistics, (iii) a calibration or sensitivity check of the response matrices, and (iv) a more careful treatment of the c1 trade-off. I do not think rejection is warranted, but the manuscript needs substantial additional experiments and a re-framing of what E2 can and cannot demonstrate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll give you the short version first: the Onion architecture is a genuinely useful wrapper for neural line-integral tomography, and the PI-injection and Softplus improvements are plausible and worth trying. But the paper's headline claim that the physics-informed loss (PILF) reduces inversion-algorithm errors is not supported. The E2 drop in Table 5 is forced—loss2 in Eq. (3-5) penalizes exactly what E2 in Eq. (4-2) measures—and the same rows show E1 getting substantially worse (from 0.50 to 1.35e-2 on Exp_EAST for VggOnion_PI, and from 0.21 to 1.94e-2 on Exp_HL-2A). Claiming that this corrects inversion errors requires either an independent ground truth or a validated forward model. Neither is provided. The response matrices are taken as given, with no calibration check; the experimental targets have known inconsistency with the inputs (that's the whole motivation); and the only datasets with true ground truth are noiseless, where loss1 alone already works. So the PILF section over-reaches.\n\nWhat is genuinely new: the Onion wrapper, the multiplication-based PI side chain, and the systematic comparison across four datasets with two backbones. The synthetic data model is simple but useful for controlled experiments. The authors are also candid about limitations: no uncertainty information, no MHD knowledge, and they promise code and data (though the repo isn't public yet).\n\nWhat's missing: error bars or multiple seeds; comparisons to published reconstruction models (Mai et al. and Wang et al. are cited but not used as baselines); out-of-device generalization tests; and an independent check on R. The per-dataset tuning of c1 also weakens the PILF claims.\n\nIf the authors reframed PILF as 'enforcing measurement consistency' rather than 'correcting inversion errors,' and added the missing evidence, the paper would be solid. As it stands, it's a conditional accept: serious referee, expected revisions. I'd bring it to reading group only if someone in the group is doing physics-informed ML for inverse problems.","headline":"Solid wrapper architecture for line-integral tomography; the PI and Softplus results hold up reasonably, but the PILF 'corrects inversion errors' claim is an artifact of the loss definition and should be reframed.","tokens_in":21675,"tokens_out":3025,"would_cite":true,"duration_ms":28244,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Physics-aware neural networks with built-in diagnostic geometry cut plasma profile reconstruction error.","keywords":["physics-informed neural networks","plasma profile reconstruction","line-integral diagnostics","soft X-ray tomography","tokamak","surrogate models","EAST","HL-2A"],"falsifier":"Take a diagnostic with a known response matrix and a validation set of independently measured ground-truth profiles (e.g., from Thomson scattering rather than from an inversion algorithm); deliberately corrupt the response matrix by rotating or shifting the lines of sight by a small amount and retrain Onion: if the corrupted-physics model matches or exceeds the clean-physics model's accuracy, or if the clean-physics reconstructions disagree with the independent measurements more than a backbone without physics, then the assumed forward model is not the source of the improvement.","tokens_in":20548,"feed_emoji":"⚛️","tokens_out":7127,"duration_ms":55329,"temperature":0.7,"pith_summary":"The paper argues that a neural network trained to invert line-integral fusion diagnostics (soft X-ray measurements on the EAST and HL-2A tokamaks) becomes substantially more accurate when the diagnostic's geometry is injected directly into the network and when a line-integral consistency term is added to the loss. The authors propose a reusable wrapper architecture, Onion, that can wrap standard backbones such as VGG and ResNet, and they report reductions in average relative reconstruction error of about $0.84\\times 10^{-2}$ on synthetic datasets and about $0.06\\times 10^{-2}$ on experimental datasets from the physics-information side chain alone, with further gains from a Softplus output activation and from the physics-informed loss. If these claims hold, surrogate models could reconstruct 2D plasma profiles in milliseconds rather than seconds, making real-time monitoring and between-shot analysis more practical for fusion reactors. The paper also introduces a synthetic data generator for line-integral diagnostics and releases experimental soft X-ray datasets, so the architecture can be tested and extended by others.","feed_headline":"Physics-aware neural nets halve plasma profile errors","feed_subtitle":"Adding line-of-sight geometry and a consistency loss makes fusion plasma reconstructions faster and more accurate.","key_machinery":"The key machinery is the Onion architecture: a physical-information (PI) side chain that takes the diagnostic response matrices $R_i$ (path lengths of each line of sight through each pixel, arranged as $(n, z, r)$ tensors), applies positional encoding and normalization, and passes them through a simplified VGG feature extractor; the resulting features are fused with the backbone's output by element-wise multiplication, so that the network's activations are multiplicatively gated by where the diagnostic 'looks.' The second component is the physics-informed loss function (PILF), $L = w_1 \\mathrm{loss}_1 + w_2 \\mathrm{loss}_2 + \\lambda L_2$, where $\\mathrm{loss}_1$ is the MSE between reconstruction $Net(x)$ and target profile $y$, and $\\mathrm{loss}_2 = \\frac{1}{n}\\sum_j (R_j \\cdot Net(x) - x_j)^2$ enforces that the back-projection of the reconstruction matches the measured chords; the weight $w_2 = c_1 \\cdot \\mathrm{loss}_1/\\mathrm{loss}_2$ dynamically balances the two terms. A third, smaller component is the use of Softplus (instead of ReLU) in the final two fully connected layers, which smooths edge artifacts in the reconstructed profiles. The response matrices are the load-bearing physics input: they encode the line-integration principle, Eq. (2-1): $x_i = R_i \\cdot y + \\Delta_i$.","core_discovery":"The central claim is that the Onion architecture—a wrapper that fuses a positional-encoded response-matrix side chain with a backbone network through element-wise multiplication and augments the loss with a back-projection consistency term—improves the accuracy of neural-network reconstruction of 2D plasma emissivity profiles from line-integral measurements. On synthetic data with essentially zero forward-model error, adding the physics information lowered the average relative error $E_1$ by roughly $0.84\\times 10^{-2}$ (about 52%), and on experimental data it lowered $E_1$ slightly while leaving the back-projection error $E_2$ almost unchanged; the physics-informed loss component then reduced $E_2$ by up to $4.59\\times 10^{-2}$ on the HL-2A dataset, at the cost of larger $E_1$, because the loss pulls predictions toward agreement with the raw diagnostic inputs rather than toward the (error-containing) inversion-algorithm targets.","pith_inferences":["If response matrices for other line-integral diagnostics (bolometry, interferometry, neutron cameras) are known, the same Onion wrapper should transfer directly, making it a candidate standard module for diagnostic surrogates.","The trade-off between $E_1$ and $E_2$ controlled by the hyperparameter $c_1$ suggests a tunable operating point: for physics-faithful reconstructions one may accept larger deviation from inversion-algorithm labels, and an independent calibration dataset (e.g., from Thomson scattering) could select the optimal $c_1$.","The lack of uncertainty information in the experimental datasets is a stated limitation; an immediate testable extension is to train Onion on synthetic data with injected noise and uncertainty weights to see whether the physics-side-chain benefits persist under realistic noise levels.","The Softplus gain on synthetic data ($E_1$ down about 71%) versus experimental data (about 27%) hints that edge smoothness is closely tied to noise-free labels; adding an edge-regularization term, which the authors mention as future work, might recover some of the experimental gap."],"forward_implications":["If the architecture works as claimed, the same wrapper can be reused across fusion devices and across diagnostic types whose forward model is a known response matrix, since the backbone is interchangeable and the physics input is device-specific geometry.","Reconstruction times of a few milliseconds on GPU (and under about 60 ms on CPU) versus seconds to minutes for Fourier–Bessel or nonstationary Gaussian process tomography could enable real-time plasma control and high-throughput between-shot analysis.","The physics-informed loss can partially correct errors baked into inversion-algorithm-generated training labels, because it pulls the reconstruction's back-projection toward the raw measured signals rather than only toward the target profiles.","On noise-free synthetic data, the gains from physics injection are large ($E_1$ down by roughly 52–71%), but on experimental data the gains shrink to about 15–27%, suggesting that dataset noise (the inherent $\\bar{\\varepsilon}$ error) is the main limiter of further improvement."],"supporting_citations":[{"why":"supplies the Exp_EAST dataset: 92 soft X-ray measurements with Fourier–Bessel target profiles.","marker":"[26]"},{"why":"supplies the Exp_HL-2A dataset: 40 viewing chords with NSGPT emissivity profile targets.","marker":"[25]"},{"why":"the Fourier–Bessel expansion tomographic code used to generate the EAST 2D SXR target labels.","marker":"[28,29]"},{"why":"the Bayesian nonstationary Gaussian process tomography code that produces the HL-2A target emissivity profiles.","marker":"[37]"},{"why":"introduces the physics-informed neural network concept that motivates adding a physical constraint term to the loss.","marker":"[19]"},{"why":"defines the VGG backbone used as the core of VggOnion.","marker":"[32]"},{"why":"defines the ResNet backbone used as the core of ResOnion.","marker":"[34]"},{"why":"provides the idea of injecting a physical state into a neural network, the basis of the PI side chain.","marker":"[41]"},{"why":"defines the response matrix as the path length of each line of sight through the pixels, the physical information injected.","marker":"[33]"},{"why":"uses soft X-ray back-projections for diagnostic model selection and motivates the back-projection terminology in the loss.","marker":"[31]"}],"fun_headline_variants":["Physics-aware neural nets cut plasma profile errors","Onion model embeds physics to improve fusion diagnostics","Line-integral physics-informed deep learning reduces errors","Back-projection consistency loss sharpens plasma reconstructions","Fusion plasma profiles reconstructed by physics-informed AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the response matrices $R_i$ exactly describe the diagnostic's line integrals, with negligible noise $\\Delta_i$; if the geometry is mis-calibrated, the physics side chain and consistency loss will pull reconstructions toward the wrong measurement model, and the paper offers no independent calibration check.","fun_headline_variants_meta":{"raw":{"variants":["Physics-aware neural nets cut plasma profile errors","Onion model embeds physics to improve fusion diagnostics","Line-integral physics-informed deep learning reduces errors","Back-projection consistency loss sharpens plasma reconstructions","Fusion plasma profiles reconstructed by physics-informed AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000841,"raw_usage":{"total_tokens":3693,"prompt_tokens":1002,"completion_tokens":2691,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":2619}},"tokens_in":618,"tokens_out":2691,"duration_ms":17632,"temperature":1.0,"reasoning_tokens":2619,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:29:36.507907+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a diagnostic with a known response matrix and a validation set of independently measured ground-truth profiles (e.g., from Thomson scattering rather than from an inversion algorithm); deliberately corrupt the response matrix by rotating or shifting the lines of sight by a small amount and retrain Onion: if the corrupted-physics model matches or exceeds the clean-physics model's accuracy, or if the clean-physics reconstructions disagree with the independent measurements more than a backbone without physics, then the assumed forward model is not the source of the improvement.","supporting_citations":[{"cited_title":"& Jenko, F","cited_arxiv_id":null,"evidence_quote":"uses soft X-ray back-projections for diagnostic model selection and motivates the back-projection terminology in the loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Exp_EAST dataset: 92 soft X-ray measurements with Fourier–Bessel target profiles."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Exp_HL-2A dataset: 40 viewing chords with NSGPT emissivity profile targets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"the Bayesian nonstationary Gaussian process tomography code that produces the HL-2A target emissivity profiles."},{"cited_title":"Reconstruction of soft X -ray and tungsten concentration profiles in Tokamaks using Bayesian method","cited_arxiv_id":null,"evidence_quote":"defines the response matrix as the path length of each line of sight through the pixels, the physical information injected."}],"review_version":1}