{"id":"5bc45204-8f81-4978-99e1-aebd424b2bee","arxiv_id":"2504.19861","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Random Forest models trained on MAST-U data predict pedestal height more accurately than simple power-law scalings, and Pareto optimization identifies discharges trading off ELM-free time against normalized pressure.","lead":"HIPED, a Random Forest toolkit trained on MAST-U spherical tokamak data, reports pedestal height predictions with R^2 around 0.7 and finds that the standard pedestal width-height power law does not hold at low aspect ratio. It also uses Pareto fronts to rank discharges by trade-offs between time-to-ELM and core pressure, offering a planning tool for experiments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random per-time-slice split in Sec. V.A likely inflates RF R² via same-discharge leakage; a shot-grouped evaluation must be run before the claimed predictive accuracy is accepted.","rationale":"I agree with the reader's weakest_assumption. The strongest claim is the quantitative accuracy of the RF pedestal-height predictor, and every test-set R² in Figures 3, 5, 6, and 17 depends on the train/test split. Section V.A explicitly describes a random split of time slices; with 7481 points from 711 shots, within-shot autocorrelation makes this a near-duplicate split. This is not a disagreement with consensus physics; it is an internal validity issue with the evaluation protocol. The proposed check is standard and decisive: a grouped split by discharge would either confirm the R² or expose the leakage. If the R² survives grouped splitting, the central claim is largely rehabilitated despite the width-model overclaim; if it drops, the main quantitative conclusions and the Pareto-optimization guidance that depends on the RF models would need to be re-derived. The descriptive finding that Delta_ped versus beta_theta,ped has low R² in-sample is not affected by the split, but it is not the load-bearing claim. I therefore do not change the reader's verdict; the concern is identical and the verdict should remain REJECT until the grouped-split evidence is supplied.","tokens_in":26324,"tokens_out":3583,"duration_ms":37390,"concrete_test":"Re-run the beta_theta,ped Random Forest with a grouped split by discharge: randomly assign 30% of the 711 shots to training and the remaining 70% to testing, keeping all time slices of a shot in one partition, and use the same features, hyperparameters, and approximate test fraction. Report test R² for the full-feature model, the no-beta_N model, the control-room-only model, and the n_e,ped and T_e,ped models. If the grouped-split R² falls substantially (e.g., below 0.6 for beta_theta,ped), the reported accuracy is inflated by autocorrelation leakage. As a companion check, evaluate the linear and power-law baselines from Figure 2 on the same grouped test split so the RF-versus-linear comparison is fair, and report the effective number of independent samples per discharge.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim is that HIPED accurately predicts pedestal height, specifically beta_theta,ped with test R² ≈ 0.755 (H-mode all), 0.713 (pre-ELM), 0.753 (ELM-free) and R² ≈ 0.674 for the control-room-only model. The weakest assumption is the data split in Section V.A: the 7481 H-mode time slices come from only 711 shots (Section IV), yet the data are 'randomly split' 30/70 per time slice with no grouping by discharge. Consecutive time slices within a discharge are strongly autocorrelated in both input features (Ip, Pbeam, shaping, fG) and target (beta_theta,ped), so the test set contains near-duplicates of training points from the same shots. A Random Forest can effectively interpolate along each shot's trajectory, inflating test R² relative to genuine shot-to-shot generalization. The same leakage affects the control-room model, the n_e,ped and T_e,ped models, and the feature-importance/SHAP interpretation. The comparison in Figure 2 is also asymmetric: the linear/power-law fits are in-sample regressions while the RF is evaluated on a held-out test set. Taken together, the central quantitative claim of accurate pedestal-height prediction is not established as stated. A secondary issue is that the abstract claims pedestal width prediction, but no width model is trained anywhere in the paper; width appears only as an input feature and in descriptive scalings.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces HIPED, a Random Forest (RF) machine-learning framework trained on MAST-U pedestal data from the third campaign. The models predict pedestal height measures (β_θ,ped, n_e,ped, T_e,ped) from engineering and physics parameters, report test-set R² around 0.71–0.76 for β_θ,ped (0.674 with 'control room' inputs only), and show that simple power-law scalings such as Δ_ped~√β_θ,ped do not reproduce the data. The paper also applies the RF surrogates to construct synthetic Pareto fronts for multi-objective optimization, ranks actual MAST-U discharges by their proximity to those fronts, and identifies control-room settings associated with the best discharges for four weightings. The central claims are accurate pedestal-height prediction, the value of additional shaping and pressure parameters, and a practical pathway for experiment planning.","tokens_in":26577,"tokens_out":3196,"duration_ms":33122,"significance":"If the predictive accuracy claims survive proper evaluation, HIPED would be a useful, interpretable tool for MAST-U experimental planning and for comparing spherical-tokamak pedestal behavior with conventional-aspect-ratio scalings. The paper provides a clearly described database filtering procedure and a GitLab repository for the pedestal database tools, which will aid reproducibility. The feature-importance and SHAP analyses give concrete, physically interpretable hypotheses about which parameters control the MAST-U pedestal. However, the central quantitative claim—that the RF accurately predicts pedestal height on unseen data—is currently undermined by the data-splitting methodology, so the significance is conditional on re-analysis with a discharge-grouped split.","major_comments":[{"comment":"The random train/test split is performed per time slice rather than per discharge. Section IV states that the 'H-mode all' category contains 7,481 time slices from 711 shots, so consecutive time slices from the same discharge almost certainly appear in both the 30% training set and the 70% test set. Time slices within a discharge are strongly autocorrelated in both the input features (P_beam, I_p, shaping, f_G) and the target β_θ,ped, so the test set contains near-duplicates of training points. A Random Forest can effectively interpolate along each shot's trajectory, inflating the test R² values reported in Figures 3, 5, 6, and Appendix D. The authors must re-run the analysis with a shot-grouped split (e.g., split by discharge ID, or leave-one-shot-out cross-validation) and report the resulting R² values before the predictive accuracy claim can be accepted.","section":"Section V.A"},{"comment":"The comparison between the RF models and the linear/power-law regressions is asymmetric. The simple fits in Figure 2(a)–(f) are in-sample regressions on the full dataset, whereas the RF performance in Figures 3, 5, and 6 is evaluated on a held-out test set. This discrepancy by construction favors the RF (in-sample fits are not penalized for overfitting, while the RF is evaluated on unseen data). To support the statement that Random Forests are 'significantly better than linear approaches in Figure 2', the authors should evaluate the linear and power-law models on the same train/test splits (or with cross-validation) and report the out-of-sample R² for both families of models.","section":"Section IV and Figure 2"},{"comment":"The abstract and introduction claim that HIPED provides 'accurate estimates of pedestal height and width', and the introduction states that the framework 'predicts pedestal parameters such as ... width'. However, no model is trained to predict the pedestal width: Δ_ped, Δ_n_e,ped, and Δ_T_e,ped appear only as input features to the β_θ,ped models (Table III and Section V.B). The absence of width prediction models means the title's and abstract's 'width' claim is not supported by any result in the paper. The authors should either train and evaluate explicit width-prediction models or revise the claims to state that the framework predicts pedestal height using the width as an input.","section":"Abstract and Section I"},{"comment":"The Pareto fronts are constructed from RF surrogate predictions on 50,000 synthetic samples 'based on the underlying MAST-U database' (Section VI.A), and the RF models for the Pareto objectives (f_G, β_N, T_ELM, ⟨n_e⟩_L) are trained on the same database with the random time-slice split. The resulting front is thereby partly an artifact of the training-data distribution and the surrogate model's extrapolation errors. For the optimization results to be useful, the authors should validate the surrogate predictions on held-out shots, and ideally compare the predicted Pareto-optimal points with actual experimental outcomes (e.g., whether the identified discharges indeed have the claimed T_ELM, f_G, and β_N). The time-window average in Eq. (8) with T_window = 0.05 s is also a free parameter; its influence on the selected optimal discharges should be discussed or shown to be robust.","section":"Section VI and Appendix C"}],"minor_comments":[{"comment":"The abstract and Section V.C refer to 'control room parameters' but f_G is not directly controllable as it depends on the achieved density; the text does note 'We keep the f_G parameter to approximate fueling control', but this approximation should be stated earlier and more prominently.","section":"Section IV"},{"comment":"The caption says 'Pbeam is the total neutral beam bower'; 'bower' should be 'power'.","section":"Figure 1 caption"},{"comment":"The feature-selection step 'Features with Pearson correlation > 0.85 are removed' is not documented in detail: the list of removed features and the thresholds used are not provided, which affects reproducibility.","section":"Section V.A"},{"comment":"The statement 'Solutions that ignore T_ELM will achieve higher β_N, f_G, and ⟨n_e⟩_L' is presented as a finding, but it is a direct consequence of removing an objective from the Pareto optimization; rephrasing as a definitional expectation would be clearer.","section":"Section VI.B"},{"comment":"The hyperparameter description says 'maximum depth of 30 to mitigate overfitting', but a maximum depth of 30 is generally not an effective overfitting mitigation for random forests; the authors likely rely on bagging and feature subsampling. This sentence could be clarified.","section":"Appendix A"},{"comment":"Reference [25] is cited as 'In Preparation'; while acceptable in a preprint, the authors should update it to a published or archived version when available.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The data-splitting issue is the key technical concern and is fixable within the manuscript's scope: a shot-grouped evaluation would either confirm or refute the central predictive claim. The paper's abstract also overstates the width-prediction capability. The authors should be encouraged to perform the re-analysis and to make the HIPED code publicly available as promised, since the database tools are already on GitLab but the ML code is not yet released."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Paris and colleagues have put together a genuinely useful MAST-U pedestal database and shown something real: on this low-aspect-ratio dataset, the standard delta_ped ~ sqrt(beta_theta,ped) scaling essentially falls apart (R² < 0.13). That empirical result is worth having, and the HIPED framework—Random Forest plus SHAP plus Pareto optimization over control-room parameters—is a sensible way to give experimentalists actionable targets. The paper is clearly written, the pedestal fits are careful, and they include useful ablations (e.g., dropping beta_N, acknowledging the TELM model's trouble at low values). They also point to a GitLab repo for the database tools, which is a plus.\n\nThe soft spots are real and load-bearing. Section V.A splits the data randomly per time slice, not per discharge. With 7,481 H-mode slices from only 711 shots, autocorrelated samples from the same shot are almost certainly in both training and test sets. That means the R² values around 0.71–0.76 for beta_theta,ped are inflated; a Random Forest can interpolate along each shot's trajectory instead of generalizing to new shots. The same leakage affects the ne_ped, Te_ped, control-room, and SHAP results. The comparison against linear baselines is also asymmetric: the power-law fits in Figure 2 are in-sample, while the RF is evaluated on a held-out test set. You need both evaluated on the same shot-grouped split before the predictive claims stand. There's also a mismatch between the abstract, which promises pedestal width prediction, and the content: no width model is trained anywhere; width appears only as an input feature and in the descriptive scaling. For a paper whose title is \"HeIght and width Predictor,\" that is a notable gap.\n\nThe Pareto optimization section is more robust to these issues, since it is explicitly a framework for ranking discharges rather than a generalization claim, though the synthetic samples drawn from the same database make it somewhat circular as an \"optimization.\" That said, the practical guidance (which control-room parameters give ELM-free operation versus high beta_N) is plausible and useful.\n\nWho is this for? Fusion experimentalists on MAST-U and other spherical tokamaks, and anyone working on ML for pedestal prediction. The qualitative scaling result and the framework deserve a serious referee. The predictive accuracy claims, as stated, do not yet. I would send it to peer review with a clear instruction: redo the evaluation with a shot-grouped split, compare baselines fairly on the same test set, and fix the abstract to match what was actually trained. Then this could be a solid contribution.","headline":"Useful MAST-U pedestal database study with a real scaling result, but the headline predictive accuracy is not yet established because the train/test split ignores discharge grouping.","tokens_in":27159,"tokens_out":2113,"would_cite":false,"duration_ms":24759,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Random-forest models predict spherical-tokamak pedestal height at 76 percent variance explained.","keywords":["pedestal prediction","random forest","spherical tokamak","H-mode","edge-localized modes","Pareto optimization","MAST-U","control room parameters"],"falsifier":"Retrain the same random-forest models with a grouped split that keeps every time slice of a discharge in either training or testing; if the held-out $R^2$ for $\\beta_{\\theta,\\mathrm{ped}}$ falls toward the 0.62 level of the power-law regressions rather than the claimed 0.76, the accuracy claim is refuted. A second check is to test the trained models on a completely new experimental campaign without any retraining.","tokens_in":26117,"feed_emoji":"⚛️","tokens_out":9505,"duration_ms":88064,"temperature":0.7,"pith_summary":"This paper introduces HIPED, a machine-learning framework that predicts and optimizes the edge pedestal in a spherical tokamak using thousands of H-mode time slices. Its central claim is that the standard power-law scaling used on conventional tokamaks—pedestal width proportional to the square root of normalized pedestal pressure—does not hold at low aspect ratio, where a random-forest model using normalized plasma pressure, elongation, and Greenwald fraction predicts pedestal height with $R^2\\approx 0.71$–$0.76$. The paper also shows that a model trained on control-room parameters alone reaches $R^2\\approx 0.674$, and that Pareto optimization over the fitted models identifies real discharges and the control settings that best trade competing objectives such as time to the next edge-localized mode and normalized plasma pressure. If the framework is sound, it gives experiments a practical way to estimate pedestal performance from knobs they can actually turn and to plan discharges against explicit priorities.","feed_headline":"Random forests predict spherical-tokamak pedestal height at 76 percent","feed_subtitle":"A control-room-only model still explains 67 percent of variance and Pareto fronts map ELM trade-offs.","key_machinery":"The machinery is a Random Forest ensemble—hundreds of decision trees trained on bootstrapped data subsets, whose averaged output captures nonlinear interactions among many inputs—combined with SHAP feature attribution to say which drivers matter. The pedestal labels come from hyperbolic-tangent fits to density and temperature profiles within a Bayesian multi-diagnostic inference system. For optimization, the paper samples the input space synthetically, identifies Pareto-front points across four objectives, and scores real discharges with a weighted sum $S_j=\\sum_i w_i \\hat{s}_{ij}$, time-averaged over a 0.05 s window to favor sustained performance.","core_discovery":"Using data from the third MAST-U campaign, the paper fits electron density and temperature pedestals with hyperbolic-tangent profiles and finds that fitting pedestal width $\\Delta_{\\mathrm{ped}}$ to normalized pedestal pressure $\\beta_{\\theta,\\mathrm{ped}}$ gives $R^2 < 0.13$, so the conventional $\\Delta_{\\mathrm{ped}}\\propto\\sqrt{\\beta_{\\theta,\\mathrm{ped}}}$ scaling fails at low aspect ratio. A random-forest regressor predicts $\\beta_{\\theta,\\mathrm{ped}}$ on a held-out 70 percent of the data with $R^2\\approx 0.755$ (0.713 for pre-ELM slices, 0.753 for ELM-free slices), with $\\beta_N$ the dominant feature; removing $\\beta_N$ drops $R^2$ to 0.585. Separate forests predict pedestal density with $R^2=0.64$ and pedestal temperature with $R^2=0.73$. Restricting inputs to seven control-room parameters still gives $R^2\\approx 0.674$. The same models are used to generate 50,000 synthetic points, build Pareto fronts over normalized pressure, Greenwald fraction, line-averaged density, and time-to-next-ELM, and rank actual discharges with a weighted Multi-Criteria Decision-Making score, yielding the control-room trajectories that approach each Pareto-optimal regime.","pith_inferences":["The reported accuracy probably overstates true generalization because the random 30/70 split lets time slices from the same discharge appear in both training and testing; a shot-grouped split would give the honest number.","A testable next step would be to use the random-forest model as an emulator for symbolic regression, distilling the low-aspect-ratio width–height relation into a compact scaling law that the paper does not provide.","The Pareto-optimal control-room prescriptions could be turned into a prospective experimental test: set the identified current, shaping, and beam powers on a new shot and check whether the predicted performance and ELM timing are realized.","Because the paper notes the random-forest model struggles at very small time-to-ELM values, Pareto fronts that rely on short ELM intervals should be treated as the least trustworthy region of the optimization."],"forward_implications":["At low aspect ratio, pedestal width cannot be treated as a single-variable function of pedestal pressure; models that include global pressure, elongation, and Greenwald fraction are needed.","Operators can estimate pedestal height to about 67 percent variance explained using only control-room quantities, so real-time or near-real-time guidance is plausible without full profile reconstruction.","Pareto-optimal discharges that maximize time to the next ELM systematically pay for it with lower normalized pressure and Greenwald fraction, while prioritizing pressure tends to require lower elongation and higher south-beam power.","The framework's ranking and local derivative of the MCDM score indicate which parameters to change next to a high-performing discharge, pointing to concrete experiment design choices.","Because the method is data-driven and device-agnostic, the same pipeline can be retrained on later campaigns or other tokamaks to produce analogous predictors and Pareto fronts."],"supporting_citations":[{"why":"supplies the automated tanh pedestal fits that define the height and width labels in the database","marker":"[25]"},{"why":"states the width–height scaling that the paper tests and finds inadequate at low aspect ratio","marker":"[58]"},{"why":"originates the conventional tokamak width–height scaling that fails on the spherical-tokamak data","marker":"[57]"},{"why":"provides the prior random-forest pedestal prediction approach on a conventional tokamak that HIPED adapts to low aspect ratio","marker":"[52]"},{"why":"describes the spherical tokamak device whose third campaign supplies the dataset","marker":"[22]"},{"why":"describes the Bayesian multi-diagnostic inference system used to construct the pedestal fits","marker":"[24]"},{"why":"is the random-forest method underlying all predictive models","marker":"[63]"},{"why":"defines the Greenwald fraction used as a key feature and optimization objective","marker":"[61]"}],"fun_headline_variants":["HIPED: ML framework scores 76% R² on spherical tokamak pedestal height","Spherical tokamak pedestals: ML beats power-law scaling with 76% R²","Random forest predicts tokamak pedestal height with 76% explained variance","ML-based HIPED predicts spherical tokamak pedestal height at R²=0.76","HIPED: ML predicts and optimizes spherical tokamak pedestals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the 7,481 H-mode time slices are independent samples, because the 30/70 train/test split is random across time slices rather than grouped by discharge, and slices from the same shot therefore appear on both sides of the split and can inflate the reported $R^2$ values.","fun_headline_variants_meta":{"raw":{"variants":["HIPED: ML framework scores 76% R² on spherical tokamak pedestal height","Spherical tokamak pedestals: ML beats power-law scaling with 76% R²","Random forest predicts tokamak pedestal height with 76% explained variance","ML-based HIPED predicts spherical tokamak pedestal height at R²=0.76","HIPED: ML predicts and optimizes spherical tokamak pedestals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001538,"raw_usage":{"total_tokens":6203,"prompt_tokens":1043,"completion_tokens":5160,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":5048}},"tokens_in":659,"tokens_out":5160,"duration_ms":33679,"temperature":1.0,"reasoning_tokens":5048,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:41:32.946329+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same random-forest models with a grouped split that keeps every time slice of a discharge in either training or testing; if the held-out $R^2$ for $\\beta_{\\theta,\\mathrm{ped}}$ falls toward the 0.62 level of the power-law regressions rather than the claimed 0.76, the accuracy claim is refuted. A second check is to test the trained models on a completely new experimental campaign without any retraining.","supporting_citations":[],"review_version":1}