pith. sign in

arxiv: 2509.07024 · v3 · pith:4X2OE44Cnew · submitted 2025-09-07 · ⚛️ physics.plasm-ph · cs.LG

TGLF-WINN: Data-Efficient Deep Learning Surrogate for Turbulent Transport Modeling in Fusion

Yadi Cao , Futian Zhang , Wesley Liu , Tom Neiser , Orso Meneghini , Lawson Fuller , Sterling Smith , Raffi Nazikian
show 2 more authors
Brian Sammuli Rose Yu
This is my paper

Pith reviewed 2026-05-21 23:02 UTC · model grok-4.3

classification ⚛️ physics.plasm-ph cs.LG
keywords turbulent transportneural network surrogateBayesian active learningTGLF modelfusion plasmasdata-efficient trainingtokamak simulationswavenumber regularization
0
0 comments X

The pith

TGLF-WINN matches full-data neural network accuracy for tokamak turbulent transport using only 25 percent of the training samples.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a neural network surrogate called TGLF-WINN for the Trapped Gyro-Landau Fluid model that predicts turbulent transport fluxes in fusion plasmas. It introduces feature engineering to simplify targets, a wavenumber-resolved regularization term that applies physics constraints to per-mode predictions, and Bayesian active learning to pick the most informative training points. These changes let the surrogate reach offline accuracy close to a standard full-data network while using far less data, cutting the cost of building surrogates for expensive gyrokinetic calculations. The approach also yields a 45-fold speedup in a downstream flux-matching workflow with similar reconstruction quality.

Core claim

TGLF-WINN demonstrates that wavenumber-informed regularization combined with Bayesian active learning produces a surrogate whose predictions of transport fluxes match those of a fully trained TGLF-NN when trained on only one-quarter of the data. Feature tuning and the regularization term together reduce relative RMSLE by 12.5 percent on the full dataset and limit degradation to an order of magnitude less than TGLF-NN when data is reduced to roughly one-ninth the original size. The resulting model remains fully differentiable and supports gradient-based coupling in whole-device simulations.

What carries the argument

The wavenumber-resolved regularization term, which imposes a physics-guided constraint on per-mode fluxes to improve generalization when training data are sparse.

If this is right

  • The surrogate delivers a 45x speedup over direct TGLF evaluations in flux-matching workflows while preserving comparable reconstruction accuracy.
  • TGLF-WINN maintains accuracy within 4.3 percent of its own full-data result when trained on only 25 percent of the samples.
  • The same regularization and active-learning strategy reduces RMSLE degradation by an order of magnitude relative to TGLF-NN when training data are cut to approximately one-ninth the full set.
  • The fully differentiable surrogate enables gradient-based optimization and coupling inside larger tokamak simulation codes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same combination of wavenumber regularization and uncertainty-driven sample selection could reduce data requirements for neural surrogates of other expensive plasma models such as full gyrokinetic codes.
  • Because the method produces a differentiable surrogate, it could be embedded directly into real-time plasma control loops that adjust actuator settings based on predicted transport.
  • Extending the active-learning loop to include new experimental data from tokamaks would allow the surrogate to improve continuously without requiring a new full offline training campaign.

Load-bearing premise

The wavenumber-resolved regularization improves generalization on sparse data without introducing systematic bias into the per-mode flux predictions.

What would settle it

A comparison of TGLF-WINN per-mode flux predictions against direct TGLF calculations on a held-out set of plasma conditions, trained with 25 percent of the data, that shows larger systematic deviations than the reported 2.8 percent offline accuracy difference.

Figures

Figures reproduced from arXiv: 2509.07024 by Brian Sammuli, Futian Zhang, Lawson Fuller, Orso Meneghini, Raffi Nazikian, Rose Yu, Sterling Smith, Tom Neiser, Wesley Liu, Yadi Cao.

Figure 1
Figure 1. Figure 1: Overview of TGLF-SINN architecture. (a) TGLF-SINN takes 31 input features (the same as TGLF-NN) and splits into 24 branches. Each branch produces 4 flux predictions corresponding to a specific wavenumber. These fluxes are then summed to yield the final fluxes. (b) The Encoder-ResNet-Decoder module shared by all prediction branches. 4.1 Dataset Our dataset is generated by running the numerical solver of TGL… view at source ↗
Figure 2
Figure 2. Figure 2: Prediction accuracy of TGLF-SINN. Heatmaps compare predicted fluxes (Pred) y ′ against ground truth (GT) fluxes y for electron particle flux (Γe), ion momentum flux (Πi), electron heat flux (Qe), and ion heat flux (Qi). The diagonal line represents perfect agreement. Color intensity indicates sample density, with darker regions representing higher concentrations of data points. Our method reduces the RMSLE… view at source ↗
Figure 3
Figure 3. Figure 3: Offline error comparison across different models. Root Mean Squared Logarithmic Error (RMSLE) (×10−2 ) is used as the metric. The graph compares errors for all four flux channels and their average over the full dataset, consistent with [34, 35]. Lower values indicate superior performance. Abbreviations: FT= Feature Tuning, SR= Spectral Regularization. 5.2 Robustness against Sparsity and Noisy Data Generati… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of TGLF-NN and TGLF-SINN performance under a sparse dataset. Models are compared with and without outlier filtering. Metrics are reported in Root Mean Squared Logarithmic Error (RMSLE) (×10−2 ) for each channel. Lower values indicate better performance. −2 0 6 −2 0 6 Log|y| Γe −2 0 8 −2 0 8 Πi −2 0 8 Log|y 0| −2 0 8 Log|y| Qe −2 0 8 Log|y 0| −2 0 8 Qi 0 50 100 0 50 100 0 50 100 0 50 100 TGLF-NN … view at source ↗
Figure 5
Figure 5. Figure 5: Heatmaps comparing prediction of removed outliers. These plots compare predicted fluxes (Pred) y ′ versus ground truth (GT) fluxes y. We compare all four channels (Γe, Πi , Qe, Qi) for TGLF-NN and TGLF-SINN over outliers in the Major perturbation dataset. The diagonal line represents perfect agreement. The colorbar represents density of samples, where darker regions have more samples. 5.3 Bayesian Active L… view at source ↗
Figure 6
Figure 6. Figure 6: Bayesian Active Learning ablation studies: (a) Error convergence comparison of [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Convergence of flux-matching in L-mode and H-mode scenarios. Comparison of normalized flux mismatch across iteration steps for TGLF, TGLF-NN, and TGLF-SINN. (a) L-mode: TGLF-SINN converges after 129 iterations, TGLF-NN after 251 iterations, while TGLF fails to converge after 1000 iterations. (b) H-mode: TGLF-SINN converges after 412 iterations, TGLF-NN after 364 iterations, while TGLF fails to converge aft… view at source ↗
Figure 8
Figure 8. Figure 8: Reconstructed flux profiles in L-mode and H-mode scenarios. Comparison of TGLF-SINN, TGLF-NN, and TGLF for predicted transport fluxes (Qe, Qi , Γe, Π) over 16 ρ locations. (a) L-mode reconstructed fluxes. (b) H-mode reconstructed fluxes. TGLF profiles are omitted due to non-convergence in both scenarios. TGLF-SINN TGLF-NN Ground Truth 0.0 0.5 1.0 0 2 4 keV Te 0.0 0.5 1.0 1 2 3 keV Ti 0.0 0.5 1.0 2.0 4.0 1 … view at source ↗
Figure 9
Figure 9. Figure 9: Reconstructed core profiles in L-mode and H-mode scenarios. Comparison of electron temperature, ion temperature, electron density, and toroidal rotation frequency profiles for experimental measurements, TGLF-NN, and TGLF-SINN. (a) L-mode scenario profiles. (b) H-mode scenario profiles. TGLF is omitted due to non-convergence in both scenarios. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of measured profiles to predictions by [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of MSLE convergence errors of [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Extended dataset validation results. (a) Relative error comparison between TGLF-SINN and TGLF-NN for reconstructed profiles in H-mode and L-mode across the DIII-D dataset. Metrics are reported as mean absolute relative error for each reconstructed profile (Te, Ti , ne) as a percentage. (b) Comparison of stored thermal energy predictions. TGLF-NN is more accurate by 10 percentage points compared to the IPB… view at source ↗
Figure 13
Figure 13. Figure 13: Impact of latent dimensions on model performance. The graph shows the test RMSLE of TGLF-SINN on [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Effect of hidden layer depth on model performance. The graph illustrates the test RMSLE of TGLF [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Influence of ResNet quantity on model performance. The graph depicts the test RMSLE of TGLF-SINN on [PITH_FULL_IMAGE:figures/full_fig_p022_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Radial profiles of mean values and standard deviations for input parameters. Each subplot represents a [PITH_FULL_IMAGE:figures/full_fig_p023_16.png] view at source ↗
read the original abstract

The Trapped Gyro-Landau Fluid (TGLF) model provides fast, accurate predictions of turbulent transport in tokamaks, but whole device simulations requiring thousands of evaluations remain computationally expensive. Neural network (NN) surrogates offer accelerated inference with fully differentiable approximations that enable gradient-based coupling but typically require large training datasets to capture transport flux variations across plasma conditions, creating significant training burden and limiting applicability to expensive gyrokinetic simulations. We propose TGLF-WINN (Wavenumber-Informed Neural Network) with three key innovations: (1) principled feature engineering that reduces target prediction range, simplifying the learning task; (2) physics-guided wavenumber-resolved regularization to improve generalization under sparse data; and (3) Bayesian Active Learning (BAL) to strategically select training samples based on model uncertainty, reducing data requirements while maintaining accuracy. Feature tuning and wavenumber regularization together deliver a 12.5% relative RMSLE reduction over TGLF-NN on the full dataset; under sparse, unfiltered training (approximately 1/9 the full size) they yield an order-of-magnitude smaller RMSLE degradation than TGLF-NN, with the wavenumber-informed regularization imposing a physics-guided constraint on per-mode fluxes. Adding Bayesian Active Learning, TGLF-WINN matches TGLF-NN's full-data offline accuracy using only 25% of the training data, within 2.8% of TGLF-NN's full-data baseline and 4.3% of our own full-data result. A downstream flux-matching workflow further shows practicality: the NN surrogate gives a 45x speedup over TGLF with comparable reconstruction accuracy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces TGLF-WINN, a neural-network surrogate for the Trapped Gyro-Landau Fluid (TGLF) turbulent-transport model. It combines feature engineering that reduces target range, wavenumber-resolved regularization intended to supply a physics constraint, and Bayesian active learning to achieve accurate predictions with substantially reduced training data. Reported results include a 12.5 % RMSLE improvement over TGLF-NN on the full dataset, an order-of-magnitude smaller degradation on sparse unfiltered data, and, with BAL, matching TGLF-NN full-data accuracy using only 25 % of the training set while delivering a 45× speedup in a downstream flux-matching workflow.

Significance. If the accuracy and data-efficiency claims hold under rigorous validation, the approach would materially lower the training burden for differentiable surrogates of expensive plasma-transport models, enabling their routine use inside whole-device simulations and gradient-based optimization loops. The explicit combination of physics-guided regularization with uncertainty-driven sample selection is a concrete contribution to the data-scarcity problem that limits surrogate modeling in fusion research.

major comments (2)
  1. [Abstract / Results] Abstract, innovations paragraph 2 and associated results section: the assertion that wavenumber-resolved regularization 'imposes a physics-guided constraint on per-mode fluxes' without systematic bias rests on aggregate RMSLE values. No mode-resolved signed-error distributions or flux-spectrum comparisons versus the TGLF reference are shown on the sparse splits; such diagnostics are required to confirm that the regularization does not shift the predicted k-spectrum even while lowering scalar error.
  2. [Results] Data-efficiency results (25 % data claim): the manuscript reports concrete RMSLE numbers and speedups but provides no description of train/validation/test splits, hyperparameter-search protocol, or statistical significance tests for the reported differences. These omissions leave the central data-reduction claim only moderately supported.
minor comments (2)
  1. [Abstract] Clarify the precise relationship between the 'approximately 1/9 the full size' unfiltered sparse set and the 25 % BAL-selected set; a table or explicit statement would remove ambiguity.
  2. [Methods] The regularization coefficient and the exact functional form of the wavenumber term should be stated explicitly (including any dependence on local plasma parameters) so that the method can be reproduced.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive review and for recognizing the potential significance of TGLF-WINN for reducing training burdens in plasma-transport surrogate modeling. We address the two major comments below and will revise the manuscript to incorporate additional diagnostics and experimental details as requested.

read point-by-point responses

  1. Referee: [Abstract / Results] Abstract, innovations paragraph 2 and associated results section: the assertion that wavenumber-resolved regularization 'imposes a physics-guided constraint on per-mode fluxes' without systematic bias rests on aggregate RMSLE values. No mode-resolved signed-error distributions or flux-spectrum comparisons versus the TGLF reference are shown on the sparse splits; such diagnostics are required to confirm that the regularization does not shift the predicted k-spectrum even while lowering scalar error.

    Authors: We agree that the current presentation relies primarily on aggregate RMSLE to support the claim of a physics-guided constraint without systematic bias. To strengthen this, the revised manuscript will include mode-resolved signed-error distributions and direct comparisons of the predicted versus reference flux spectra (across wavenumber bins) specifically for the sparse training splits. These additions will explicitly verify that the regularization preserves the k-spectrum shape and does not introduce per-mode shifts, thereby providing the requested confirmation. revision: yes

  2. Referee: [Results] Data-efficiency results (25 % data claim): the manuscript reports concrete RMSLE numbers and speedups but provides no description of train/validation/test splits, hyperparameter-search protocol, or statistical significance tests for the reported differences. These omissions leave the central data-reduction claim only moderately supported.

    Authors: We acknowledge that the manuscript currently lacks explicit documentation of these experimental details. In the revised version we will add a dedicated subsection describing the train/validation/test split construction (including how the 25% subset was selected), the full hyperparameter-search protocol (search space, optimization method, and final choices), and statistical significance testing (e.g., results from multiple random seeds with error bars or paired statistical tests on the RMSLE differences). These additions will make the data-efficiency results more rigorously supported. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML surrogate validated against external TGLF baselines

full rationale

The paper reports empirical accuracy and data-efficiency results for TGLF-WINN via direct comparisons of RMSLE and reconstruction error against TGLF and TGLF-NN on held-out plasma data. Feature engineering, wavenumber regularization, and Bayesian active learning are implemented as standard training components whose effects are measured by performance deltas rather than by any equation that reduces a reported prediction to a fitted input by construction. No self-citation load-bearing steps, uniqueness theorems, or ansatz smuggling appear in the derivation of the central claims; the 25% data result and 12.5% RMSLE improvement are outcomes of experimental splits and hyperparameter choices, not tautological redefinitions of the inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim depends on TGLF serving as reliable ground truth, standard neural-network optimization assumptions, and several untuned or lightly tuned hyperparameters for regularization and active learning.

free parameters (2)
  • Wavenumber regularization coefficient
    Strength of the physics-guided per-mode penalty term must be chosen or cross-validated.
  • BAL acquisition function hyperparameters
    Parameters controlling uncertainty-based sample selection are set by the authors.
axioms (1)
  • domain assumption TGLF model outputs constitute accurate enough targets for surrogate training across the sampled plasma conditions
    All training and evaluation treat TGLF fluxes as ground truth.

pith-pipeline@v0.9.0 · 5871 in / 1347 out tokens · 66350 ms · 2026-05-21T23:02:38.443243+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

58 extracted references · 58 canonical work pages

  1. [1]

    Bonoli, L

    P. Bonoli, L. C. McInnes, C. Sovinec, D. Brennan, T. Rognlien, P. Snyder, J. Candy, C. Kessel, J. Hittinger, L. Chacon, and . others. Report of the workshop on integrated simulations for magnetic fusion energy sciences. Office of Fusion Energy Sci. and the Office of Adv. Sci. Comput. Res., Tech. Rep, 2015

    work page 2015

  2. [2]

    Suchyta, S

    E. Suchyta, S. Klasky, N. Podhorszki, M. Wolf, A. Adesoji, C. Chang, J. Choi, P. E. Davis, J. Dominski, S. Ethier, and . others. The exascale framework for high fidelity coupled simulations (effis): Enabling whole device modeling in fusion science.The International Journal of High Performance Computing Applications, 36(1):106–128, 2022

    work page 2022

  3. [3]

    Meneghini, T

    O. Meneghini, T. Slendebroek, B.C. Lyons, K. McLaughlin, J. McClenaghan, L. Stagner, J. Harvey, T.F. Neiser, A. Ghiozzi, G. Dose, J. Guterl, A. Zalzali, T. Cote, N. Shi, D. Weisberg, S.P. Smith, B.A. Grierson, and J. Candy. FUSE (Fusion Synthesis Engine): A Next Generation Framework for Integrated Design of Fusion Pilot Plants. arXiv, 2024

    work page 2024

  4. [4]

    Hirshman and D

    S. Hirshman and D. Sigmar. Neoclassical transport of impurities in tokamak plasmas.Nuclear Fusion, 21(9):1079, 1981

    work page 1981

  5. [5]

    E. A. Belli and J. Candy. Full linearized fokker–planck collisions in neoclassical transport simulations.Plasma Physics and Controlled Fusion, 54(1):015015, dec 2011

    work page 2011

  6. [6]

    Snyder, R

    P. Snyder, R. Groebner, A. Leonard, T. Osborne, and H. Wilson. Development and validation of a predictive model for the pedestal height.Physics of Plasmas, 16(5), 2009

    work page 2009

  7. [7]

    Snyder, R

    P. Snyder, R. Groebner, J. Hughes, T. Osborne, M. Beurskens, A. Leonard, H. Wilson, and X. Xu. A first-principles predictive model of the pedestal height and width: development, testing and iter optimization with the eped model. Nuclear Fusion, 51(10):103016, 2011

    work page 2011

  8. [8]

    L. Lao, K. Burrell, T. Casper, V . Chan, M. Chu, J. DeBoo, E. Doyle, R. Durst, C. Forest, C. Greenfield, and . others. Rotational and magnetic shear stabilization of magnetohydrodynamic modes and turbulence in diii-d high performance discharges.Physics of plasmas, 3(5):1951–1958, 1996

    work page 1951

  9. [9]

    B. C. Lyons, J. McClenaghan, T. Slendebroek, O. Meneghini, T. F. Neiser, S. P. Smith, D. B. Weisberg, E. A. Belli, J. Candy, J. M. Hanson, L. L. Lao, N. C. Logan, S. Saarelma, O. Sauter, P. B. Snyder, G. M. Staebler, K. E. Thome, and A. D. Turnbull. Flexible, integrated modeling of tokamak stability, transport, equilibrium, and pedestal physics.Physics of...

    work page 2023

  10. [10]

    Slendebroek, J

    T. Slendebroek, J. McClenaghan, O. M. Meneghini, B. C. Lyons, S. P. Smith, T. F. Neiser, N. Shi, and J. Candy. Elevating zero dimensional global scaling predictions to self-consistent theory-based simulations.Physics of Plasmas, 30(7):072511, 07 2023

    work page 2023

  11. [11]

    McClenaghan, A

    J. McClenaghan, A. Marinoni, A. O. Nelson, T. Neiser, L. L. Lao, G. M. Staebler, S. P. Smith, O. M. Meneghini, B. C. Lyons, P. B. Snyder, and M. Austin. Examining transport and integrated modeling predictive capabilities for negative-triangularity scenarios.Plasma Physics and Controlled Fusion, 66(11):115008, sep 2024

    work page 2024

  12. [12]

    Dimits, B

    A. Dimits, B. Cohen, N. Mattor, W. Nevins, D. Shumaker, S. Parker, and C. Kim. Simulation of ion temperature gradient turbulence in tokamaks.Nuclear fusion, 40(3Y):661, 2000

    work page 2000

  13. [13]

    Candy and R

    J. Candy and R. Waltz. An eulerian gyrokinetic-maxwell solver.Journal of Computational Physics, 186(2):545– 581, 2003

    work page 2003

  14. [14]

    Howard, C

    N. Howard, C. Holland, A. White, M. Greenwald, and J. Candy. Multi-scale gyrokinetic simulation of tokamak plasmas: enhanced heat loss due to cross-scale coupling of plasma turbulence.Nuclear Fusion, 56(1):014004, 2015

    work page 2015

  15. [15]

    Candy, E

    J. Candy, E. A. Belli, and R. Bravenec. A high-accuracy eulerian gyrokinetic solver for collisional plasmas. Journal of Computational Physics, 324:73–93, 2016. 14

    work page 2016

  16. [16]

    T. F. Neiser, F. Jenko, T. A. Carter, L. Schmitz, D. Told, G. Merlo, A. Bañón Navarro, P. C. Crandall, G. R. McKee, and Z. Yan. Gyrokinetic GENE simulations of DIII-D near-edge L-mode plasmas.Physics of Plasmas, 26(9):092510, 2019

    work page 2019

  17. [17]

    Waltz, G

    R. Waltz, G. Staebler, W. Dorland, G. Hammett, M. Kotschenreuther, and J. Konings. A gyro-landau-fluid transport model.Physics of Plasmas, 4(7):2482–2496, 1997

    work page 1997

  18. [18]

    Candy, C

    J. Candy, C. Holland, R. Waltz, M. R. Fahey, and E. Belli. Tokamak profile prediction using direct gyrokinetic and neoclassical simulation.Physics of Plasmas, 16(6), 2009

    work page 2009

  19. [19]

    Rafiq, A

    T. Rafiq, A. Kritz, J. Weiland, A. Pankin, and L. Luo. Physics basis of multi-mode anomalous transport module. Physics of Plasmas, 20(3), 2013

    work page 2013

  20. [20]

    Staebler, J

    G. Staebler, J. Kinsey, and R. Waltz. A theory-based transport model with comprehensive physics.Physics of Plasmas, 14(5), 2007

    work page 2007

  21. [21]

    Staebler, J

    G. Staebler, J. Kinsey, and R. Waltz. Gyro-landau fluid equations for trapped and passing particles.Physics of Plasmas, 12(10), 2005

    work page 2005

  22. [22]

    Kinsey, G

    J. Kinsey, G. Staebler, and R. Waltz. The first transport code simulations using the trapped gyro-landau-fluid model.Physics of Plasmas, 15(5), 2008

    work page 2008

  23. [23]

    Pfaff, M

    T. Pfaff, M. Fortunato, A. Sanchez-Gonzalez, and P. Battaglia. Learning mesh-based simulation with graph networks. InInternational conference on learning representations, 2020

    work page 2020

  24. [24]

    D. Wu, L. Gao, X. Xiong, M. Chinazzi, A. Vespignani, Y . Ma, and R. Yu. Deepgleam: a hybrid mechanistic and deep learning model for covid-19 forecasting.arXiv preprint arXiv:2102.06684, 2021

    work page arXiv 2021

  25. [25]

    Y . Cao, M. Chai, M. Li, and C. Jiang. Efficient learning of mesh-based physical simulation with bi-stride multi-scale graph neural network. InInternational conference on machine learning, pages 3541–3558. PMLR, 2023

    work page 2023

  26. [26]

    Y . Cao, Y . Liu, L. Yang, R. Yu, H. Schaeffer, and S. Osher. Vicon: Vision in-context operator networks for multi-physics fluid dynamics prediction.arXiv preprint arXiv:2411.16063, 2024

    work page arXiv 2024

  27. [27]

    I. Char, Y . Chung, W. Neiswanger, K. Kandasamy, A. O. Nelson, M. Boyer, E. Kolemen, and J. Schneider. Offline contextual bayesian optimization.Advances in Neural Information Processing Systems, 32, 2019

    work page 2019

  28. [28]

    J. Seo, S. Kim, A. Jalalvand, R. Conlin, A. Rothstein, J. Abbate, K. Erickson, J. Wai, R. Shousha, and E. Kolemen. Avoiding fusion plasma tearing instability with deep reinforcement learning.Nature, 626(8000):746–751, 2024

    work page 2024

  29. [29]

    Abbate, R

    J. Abbate, R. Conlin, and E. Kolemen. Data-driven profile prediction for diii-d.Nuclear Fusion, 61(4):046027, 2021

    work page 2021

  30. [30]

    M. E. Fenstermacher, . DIII-D Team:, J. Abbate, S. Abe, T. Abrams, M. Adams, B. Adamson, N. Aiba, T. Akiyama, P. Aleynikov, E. Allen, S. Allen, H. Anand, J. Anderson, Y . Andrew, T. Andrews, D. Appelt, R. Arbon, N. Ashikawa, A. Ashourvan, M. Aslin, Y . Asnis, M. Austin, D. Ayala, J. Bak, I. Bandyopadhyay, S. Banerjee, K. Barada, L. Bardoczi, J. Barr, E. B...

    work page 2022

  31. [31]

    Gopakumar, S

    V . Gopakumar, S. Pamela, L. Zanisi, Z. Li, A. Gray, D. Brennand, N. Bhatia, G. Stathopoulos, M. Kusner, M. P. Deisenroth, and . others. Plasma surrogate modelling using fourier neural operators.Nuclear Fusion, 64(5):056025, 2024

    work page 2024

  32. [32]

    M. M. Rahman, Z. Bai, J. R. King, C. R. Sovinec, X. Wei, S. Williams, and Y . Liu. Sparsified time-dependent fourier neural operators for fusion simulations.Physics of Plasmas, 31(12), 2024

    work page 2024

  33. [33]

    Citrin, S

    J. Citrin, S. Breton, F. Felici, F. Imbeaux, T. Aniel, J. Artaud, B. Baiocchi, C. Bourdelle, Y . Camenen, and J. Garcia. Real-time capable first principle based modelling of tokamak turbulent transport.Nuclear Fusion, 55(9):092001, 2015

    work page 2015

  34. [34]

    Meneghini, S

    O. Meneghini, S. P. Smith, P. B. Snyder, G. M. Staebler, J. Candy, E. Belli, L. Lao, M. Kostuk, T. Luce, T. Luda, and . others. Self-consistent core-pedestal transport simulations with neural network accelerated models.Nuclear Fusion, 57(8):086034, 2017

    work page 2017

  35. [35]

    Neiser, O

    T. Neiser, O. Meneghini, S. Smith, J. McClenaghan, D. Orozco, J. Hall, G. Staebler, E. Belli, and J. Candy. Database generation for validation of TGLF and retraining of neural network accelerated TGLF-NN. InBulletin of the American Physical Society, 2022

    work page 2022

  36. [36]

    K. L. Plassche, J. Citrin, C. Bourdelle, Y . Camenen, F. J. Casson, V . I. Dagnelie, F. Felici, A. Ho, S. Van Mulders, and J. Contributors. Fast modeling of turbulent transport in fusion plasmas using neural networks.Physics of Plasmas, 27(2), 2020

    work page 2020

  37. [37]

    Fransson, A

    E. Fransson, A. Gillgren, A. Ho, J. Borsander, O. Lindberg, W. Rieck, M. Åqvist, and P. Strand. A fast neural network surrogate model for the eigenvalues of qualikiz.Physics of Plasmas, 30(12), 2023

    work page 2023

  38. [38]

    Pavone, A

    A. Pavone, A. Merlo, S. Kwak, and J. Svensson. Machine learning and bayesian inference in nuclear fusion research: an overview.Plasma Physics and Controlled Fusion, 65(5):053001, 2023

    work page 2023

  39. [39]

    Settles.Active Learning, volume 6 ofSynthesis Lectures on Artificial Intelligence and Machine Learning

    B. Settles.Active Learning, volume 6 ofSynthesis Lectures on Artificial Intelligence and Machine Learning. Morgan & Claypool Publishers, 2012. 17

    work page 2012

  40. [40]

    Neiser, O

    T. Neiser, O. Meneghini, S. Smith, J. McClenaghan, T. Slendebroek, D. Orozco, B. Sammuli, G. Staebler, J. Hall, E. Belli, and J. Candy. Multi-fidelity neural network representation of gyrokinetic turbulence. InAPS Division of Plasma Physics Meeting Abstracts, volume 2023 ofAPS Meeting Abstracts, page PP11.039, January 2023

    work page 2023

  41. [41]

    Waltz and R

    R. Waltz and R. Miller. Ion temperature gradient turbulence simulations and plasma flux surface shape.Physics of Plasmas, 6(11):4265–4271, 1999

    work page 1999

  42. [42]

    Kinsey, G

    J. Kinsey, G. Staebler, and R. Waltz. Predicting core and edge transport barriers in tokamaks using the glf23 drift-wave transport model.Physics of Plasmas, 12(5), 2005

    work page 2005

  43. [43]

    G. M. Staebler, J. Candy, E. A. Belli, J. E. Kinsey, N. Bonanomi, and B. Patel. Geometry dependence of the fluctuation intensity in gyrokinetic turbulence.Plasma Physics and Controlled Fusion, 63(1):015013, nov 2020

    work page 2020

  44. [44]

    Staebler, E

    G.M. Staebler, E. A. Belli, J. Candy, J.E. Kinsey, H. Dudding, and B. Patel. Verification of a quasi-linear model for gyrokinetic turbulent transport.Nuclear Fusion, 61(11):116007, sep 2021

    work page 2021

  45. [45]

    H. G. Dudding, F. Casson, D. Dickinson, B. Patel, C. Roach, E. Belli, and G. Staebler. A new quasilinear saturation rule for tokamak turbulence with application to the isotope scaling of transport.Nuclear Fusion, 62(9):096005, 2022

    work page 2022

  46. [46]

    Staebler, J.M

    G.M. Staebler, J.M. Park, E. Hassan, C. Angioni, E. Fable, C. Bourdelle, J.E. Kinsey, C. Holland, E.A. Belli, T. Neiser, J. Candy, and R.E. Waltz. Successful prediction of tokamak transport in the l-mode regime.Nuclear Fusion, 64(8):085002, jul 2024

    work page 2024

  47. [47]

    Fischer, C

    R. Fischer, C. Wendland, A. Dinklage, S. Gori, V . Dose, W. team, and . others. Thomson scattering analysis with the bayesian probability theory.Plasma physics and controlled fusion, 44(8):1501, 2002

    work page 2002

  48. [48]

    Svensson, A

    J. Svensson, A. Dinklage, J. Geiger, and R. Fischer. An integrated data analysis model for the w7-as stellarator. In 30th EPS Conference on Contr. Fusion and Plasma Phys., St. Petersburg, volume 27, 2003

    work page 2003

  49. [49]

    Svensson, A

    J. Svensson, A. Dinklage, J. Geiger, A. Werner, and R. Fischer. Integrating diagnostic data analysis for w 7-as using bayesian graphical models.Review of Scientific Instruments, 75(10):4219–4221, 2004

    work page 2004

  50. [50]

    Svensson, A

    J. Svensson, A. Werner, J. Contributors, and . others. Current tomography for axisymmetric plasmas.Plasma Physics and Controlled Fusion, 50(8):085002, 2008

    work page 2008

  51. [51]

    O. P. Ford.Tokamak plasma analysis through Bayesian diagnostic modelling. PhD thesis, Imperial College London, 2010

    work page 2010

  52. [52]

    S. Kwak, J. Svensson, S. Bozhenkov, J. Flanagan, M. Kempenaars, A. Boboc, Y . Ghim, and J. Contributors. Bayesian modelling of thomson scattering and multichannel interferometer diagnostics using gaussian processes. Nuclear Fusion, 60(4):046009, 2020

    work page 2020

  53. [53]

    S. Kwak, J. Svensson, O. Ford, L. Appel, Y . Ghim, and J. Contributors. Bayesian inference of axisymmetric plasma equilibrium.Nuclear Fusion, 62(12):126069, 2022

    work page 2022

  54. [54]

    Rodriguez-Fernandez, N

    P. Rodriguez-Fernandez, N. Howard, A. Saltzman, S. Kantamneni, J. Candy, C. Holland, M. Balandat, S. Ament, and A. White. Enhancing predictive capabilities in fusion burning plasmas through surrogate-based optimization in core transport solvers.Nuclear Fusion, 64(7):076034, 2024

    work page 2024

  55. [55]

    J. B. Lestz, G. Avdeeva, T. F. Neiser, M. V . Gorelenkova, F. D. Halpern, S. M. Kaye, J. McClenaghan, A. Y . Pankin, and K. E. Thome. Assessing time-dependent temperature profile predictions using reduced transport models for high performing nstx plasmas, 2025

    work page 2025

  56. [56]

    Patel, P

    B. Patel, P. A. Hill, L. T. Pattinson, M. Giacomin, A. Bokshi, D. Kennedy, H. Dudding, J. F. Parisi, T. Neiser, A. Jayalekshmi, and . others. Pyrokinetics-a python library to standardise gyrokinetic analysis.Journal of open source software, 2024. 18

    work page 2024

  57. [57]

    T. F. Neiser, D. Sun, B. Agnew, T. Slendebroek, O. Meneghini, B. C. Lyons, A. G. Ghiozzi, J. T. McClenaghan, G. M. Staebler, and J. Candy. TJLF: The quasi-linear model of gyrokinetic transport TGLF translated to Julia. In Bulletin of the American Physical Society, 2024

    work page 2024

  58. [58]

    Pratt, V

    Q. Pratt, V . Hall-Chen, T.F. Neiser, R. Hong, J. Damba, T.L. Rhodes, K.E. Thome, J. Yang, S.R. Haskey, T. Cote, and T. Carter. Density wavenumber spectrum measurements, synthetic diagnostic development, and tests of quasilinear turbulence modeling in the core of electron-heated diii-d h-mode plasmas.Nuclear Fusion, 64(1):016001, nov 2023. 19 A Complete A...

    work page 2023