pith. sign in

arxiv: 2606.19523 · v1 · pith:UIJ24HVBnew · submitted 2026-06-17 · ⚛️ physics.plasm-ph

Bayesian optimization of stellarator alpha-particle confinement using data-informed parameter spaces and dimensionality reduction

Matt Landreman , Michael Czekanski , Andrew Giuliani , Byoungchan Jang , Rory Conlin This is my paper

Pith reviewed 2026-06-26 18:39 UTC · model grok-4.3

classification ⚛️ physics.plasm-ph
keywords stellarator optimizationBayesian optimizationalpha-particle confinementplasma boundary parameterizationprincipal component analysisquantile transformationfast-particle confinementdimensionality reduction
0
0 comments X

The pith

Stellarators optimized via data-informed parameter spaces achieve excellent alpha-particle confinement even far from quasisymmetry.

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

The paper develops two new ways to parameterize stellarator plasma boundaries for optimization algorithms. Starting from a collection of known stellarator shapes, it applies quantile transformations to Fourier coefficients or combines principal component analysis with quantiles to create bounded parameters between zero and one. These spaces reduce the chance of invalid self-intersecting surfaces during optimization. Bayesian optimization using these parameters, combined with particle tracing for confinement, produces designs with strong fast-particle performance. The results show that good confinement does not require the plasma to be close to quasisymmetric or quasi-isodynamic.

Core claim

By transforming Fourier amplitudes of stellarator boundaries using quantile maps from existing designs, or by applying PCA to boundary points and then quantiles, the degrees of freedom become naturally bounded. This enables efficient Bayesian optimization of alpha-particle confinement through guiding-center tracing. The resulting configurations demonstrate excellent confinement in magnetic fields that deviate substantially from quasisymmetry or quasi-isodynamicity.

What carries the argument

Quantile-transformed Fourier parameters and PCA-quantile reduced parameter spaces derived from a dataset of existing stellarator boundaries, which provide scaled and bounded variables for the optimization.

Load-bearing premise

The collection of existing stellarator boundaries used to build the transformations is representative of the valid shapes that can be encountered during optimization.

What would settle it

Finding that a significant fraction of points in the new parameter spaces produce self-intersecting boundaries or MHD equilibrium failures would indicate the method does not reliably generate usable surfaces.

Figures

Figures reproduced from arXiv: 2606.19523 by Andrew Giuliani, Byoungchan Jang, Matt Landreman, Michael Czekanski, Rory Conlin.

Figure 1
Figure 1. Figure 1: A quantile transformation is a piecewise-linear function that maps an arbitrary data distribution to a uniform distribution on [0, 1]. In the left and right panels, black dots show typical data points with a random vertical location so the points can be distinguished. low-dimensional representations of quasi-helically symmetric equilibria [17]. The PCA method in our work differs from these earlier studies … view at source ↗
Figure 2
Figure 2. Figure 2: Transformation pipeline between boundary Fourier amplitudes and PCA degrees of freedom. from Parseval’s theorem equating the norms of functions and their Fourier coefficients. Attractive features of the real-space approach are that it is clear that all real-space coordinates can be given equal weight in the PCA, and surfaces with different parameterizations can be combined in the same PCA. The PCA surface … view at source ↗
Figure 3
Figure 3. Figure 3: Fraction fus of the baseline and data-informed Fourier parameter spaces for which the MHD equilibrium converges to the required tolerance. Here, |m|max and |n|max refer to the maximum |m| and |n| values in eq (1). For the baseline, the width of the bound constraints was optimized to maximize fus. For the data-informed space, each shape parameter is randomly sampled uniformly on [umin, 1 − umin]. Values in … view at source ↗
Figure 4
Figure 4. Figure 4: Pareto fronts reflecting the trade-off between diversity of possible shapes vs. the fraction of the space where the MHD equilibrium converges, as defined in section 3.1. Each point represents a different choice of parameter space. Within each of the three shape parameterizations (colors), the different points correspond to different choices of |m|max = |n|max, umin, x0, number of principal components, and/… view at source ↗
Figure 5
Figure 5. Figure 5: Cross-sections for stellarator shapes using the data-informed Fourier method. The degrees of freedom are randomly sampled evenly over the optimization space. A wide range of plausible strongly-shaped boundaries can be represented with few self-intersections. 0 2 4 6 8 10 Number of Components 0.0 0.2 0.4 0.6 0.8 1.0 nfp 3 All nfp Cumulative Explained Variance Ratio [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Most of the variance in the boundary shapes in the dataset is explained by a small number of principal components. Fourier method. To convey the distribution of typical shapes, the degrees of freedom are randomly sampled evenly over the parameter space. In the figure, the domain of each design variable is [0.1, 0.9], which is sufficiently conservative to eliminate all self-intersections in the boundaries s… view at source ↗
Figure 7
Figure 7. Figure 7 [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Fraction fus of the PCA parameter space for which the MHD equilibrium converges to the required tolerance. Each parameter is randomly sampled uniformly on [0, 1]. Also shown is the fraction of the space for which the surfaces are free of self-intersection (purple). 0.0 0.2 0.4 0.6 0.8 1.0 Normalized minor radius 0.0 0.5 1.0 1.5 2.0 2.5 3.0 ne [m 3 ] 1e20 Density 0.0 0.2 0.4 0.6 0.8 1.0 Normalized minor rad… view at source ↗
Figure 9
Figure 9. Figure 9: Profiles from section 4.1 used for the MHD equilibria and the alpha particle birth distribution. pressure profile p = (Te + Ti)ne is then p(ρ) = (1.44 × 106 P a)(1 − 2ρ 2 + 2ρ 4 − ρ 6 − 2ρ 8 + 4ρ 10 − 2.8ρ 12 − 0.4ρ 14 + 2.4ρ 16 − 1.2ρ 18). (7) These profiles along with the ensuing DT fusion reaction rate are shown in figure 9. The pressure profile is used when computing MHD equilibria, and alpha particles… view at source ↗
Figure 10
Figure 10. Figure 10: The objective function (9) is chosen to allow early termination of the tracing for high-loss configurations, reserving computation time for the best-performing configurations. series regression on these four configurations gives a scaling a = (3.1m)/A0.38 where A is the aspect ratio. This scaling correctly fits the minor radius of these four configurations to two significant digits. For all equilibria in … view at source ↗
Figure 11
Figure 11. Figure 11: Illustration of the asynchronous algorithm, showing the optimization of configuration 2 below. For each of the four workers, corresponding to one of the GPUs on the compute node, each colored block indicates the evaluation of the objective function for one stellarator shape. All workers are kept busy, and for shapes for which the equilibrium fails or with poor confinement, little time is used. For the man… view at source ↗
Figure 12
Figure 12. Figure 12: Optimization history for configuration 2 (same as figure 11). minimizations of the objective f in eq (9), fixing the aspect ratio at 6. For configuration 1, we opti￾mize in the Fourier space, with modes |n| ≤ 1 and Garabedian m ∈ [0, 2] (corresponding to standard Fourier modes |m| ≤ 1). This space has 7 degrees of freedom, and the optimization algorithm was TURBO. Remarkably, even though the boundary shap… view at source ↗
Figure 13
Figure 13. Figure 13: The five new configurations obtained using Bayesian optimization. In the bottom row, cross-sections are shown at ϕ = 0 (red), π/8 (green), and π/4 (blue). 10 5 10 4 10 3 10 2 10 1 Time (s) 10 4 10 3 10 2 10 1 10 0 Energy Loss Fraction Configuration 1 Configuration 2 Configuration 3 Configuration 4 Configuration 5 [PITH_FULL_IMAGE:figures/full_fig_p015_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Alpha-particle energy losses for the five optimized configurations, with particles initialized throughout the volume at the local fusion rate. Losses are quite low, especially when Mercier stability is not enforced (configurations 1-3, loss < 0.06%). Even when Mercier stability is enforced (configurations 4-5) losses are ≤ 1%. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Demonstration of Mercier stability (DMerc > 0) for configurations 4 and 5. then, none of the configurations are close to being quasisymmetric, and while configurations 4-5 may be somewhat close to quasi-isodynamic, they depart from omnigenity significantly near Bmax. It is remarkable that such good confinement of alpha particles is possible when the deviations from QS and QI are so large. It is interestin… view at source ↗
Figure 16
Figure 16. Figure 16: Magnetic field strength on four flux surfaces of configuration 1, as functions of the poloidal and toroidal Boozer angles. Two field periods are shown, and the black diagonal line indicates the field direction. 0 B 0 2 B |B| [T], =0.25 9.5 10.0 0 B 0 2 B |B| [T], =0.5 9.0 9.5 10.0 10.5 0 B 0 2 B |B| [T], =0.75 9.0 9.5 10.0 10.5 11.0 11.5 0 B 0 2 B |B| [T], =1 8 9 10 11 12 [PITH_FULL_IMAGE:figures/full_fi… view at source ↗
Figure 17
Figure 17. Figure 17: Magnetic field strength on four flux surfaces of configuration 2, as functions of the poloidal and toroidal Boozer angles. Two field periods are shown, and the black diagonal line indicates the field direction. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Magnetic field strength on four flux surfaces of configuration 3, as functions of the poloidal and toroidal Boozer angles. Two field periods are shown, and the black diagonal line indicates the field direction. 0 B 0 2 B |B| [T], =0.25 5 6 7 8 9 10 11 0 B 0 2 B |B| [T], =0.5 5 6 7 8 9 10 11 0 B 0 2 B |B| [T], =0.75 5 6 7 8 9 10 11 12 0 B 0 2 B |B| [T], =1 6 7 8 9 10 11 12 [PITH_FULL_IMAGE:figures/full_fi… view at source ↗
Figure 19
Figure 19. Figure 19: Magnetic field strength on four flux surfaces of configuration 4, as functions of the poloidal and toroidal Boozer angles. Two field periods are shown, and the black diagonal line indicates the field direction. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Magnetic field strength on four flux surfaces of configuration 5, as functions of the poloidal and toroidal Boozer angles. Two field periods are shown, and the black diagonal line indicates the field direction. 10 5 10 4 10 3 10 2 10 1 Time [sec] 10 3 10 2 10 1 10 0 Fraction of alpha particles lost W7-X w/o ripple ARIES-CS Precise QA New config 1 New config 2 New config 3 New config 4 New config 5 [PITH_… view at source ↗
Figure 21
Figure 21. Figure 21: Collisionless losses of alpha particles for the new configurations and several well-known ones using the alternative scaling and initialization from [2]. Here, all configurations are scaled to the minor radius and volume-averaged B of ARIES-CS, and particles are initialized at s = 0.25. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Profiles of effective ripple for the new configurations here as well as several well-known stellarators. Even without optimization of ϵeff or optimization for QS or omnigenity, the new configurations have ϵeff < 1% in the core, better than W7-X, which is likely good enough. 1 0 1 2 3 4 5 Minimized objective 0 4 8 12 16 20 Count Dataset: nfp 3 Dataset: All nfp [PITH_FULL_IMAGE:figures/full_fig_p020_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Distribution of minimum values of the objective function for 192 independent optimizations, showing the dependence on which data were used to define the space. The results are insensitive to whether or not nfp = 1 and 2 configurations were included. objective, ∼ −1. This finding suggests that optimizations may not be overly sensitive to the choice of data, at least in some cases. Another test of sensitivi… view at source ↗
read the original abstract

Modern stellarators are typically designed by optimizing the shape of the plasma boundary surface, with the parameters taken to be Fourier amplitudes. Many promising optimization algorithms such as Bayesian methods require bound constraints on the parameters and are most efficient when each parameter is scaled similarly to the others. With the typical Fourier parameterization, it is unclear how to set these bounds: wide constraints lead to self-intersecting boundaries and frequent failures of the MHD equilibrium calculation, while tight bound constraints limit expressiveness. To address these issues, here we propose two new parameter spaces for stellarator optimization. Both begin with a dataset of existing stellarator boundaries. In the first approach, a quantile transformation is applied to each Fourier degree of freedom, mapping the data distribution to a uniform distribution on the unit interval. In the second approach, principal component analysis (PCA) is applied to points on the boundaries, followed by a quantile transformation. For both approaches, the transformed variables become the degrees of freedom, naturally bounded to [0, 1]. The PCA method has the additional benefit of dimensionality reduction, with high expressiveness for a small number of parameters. The methods are demonstrated via Bayesian optimization for good alpha-particle confinement with guiding-center tracing inside the optimization loop, using asynchronous parallelization. These optimizations yield stellarator configurations with excellent fast-particle confinement in fields that can be far from quasisymmetric or quasi-isodynamic.

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

3 major / 2 minor

Summary. The paper proposes two data-informed parameter spaces for stellarator boundary optimization: (1) per-Fourier-coefficient quantile transforms mapping existing stellarator data to [0,1], and (2) PCA on boundary points followed by quantile transforms, which also reduces dimensionality. These bounded spaces are used inside a Bayesian optimization loop that minimizes guiding-center alpha-particle losses, with asynchronous parallelization. The central claim is that the resulting configurations achieve excellent fast-particle confinement even when the fields are far from quasisymmetric or quasi-isodynamic.

Significance. If validated, the approach would address a practical bottleneck in stellarator design by supplying natural [0,1] bounds and optional dimensionality reduction without sacrificing expressiveness, while integrating guiding-center tracing directly in the loop. Credit is due for constructing the transforms from an external dataset of boundaries and for demonstrating the method on a concrete confinement objective.

major comments (3)
  1. [Abstract] Abstract: the claim that the optimizations 'yield stellarator configurations with excellent fast-particle confinement' is unsupported by any quantitative metrics (loss fractions, comparison baselines, or uncertainty estimates), which is load-bearing for the central result.
  2. [Abstract] Abstract and methods description: no quantitative report is given on the fraction of proposals that produced self-intersecting surfaces or caused VMEC (or equivalent) equilibrium failures during the Bayesian search; without this, it is impossible to confirm that the quantile/PCA transforms keep the optimizer inside the valid domain.
  3. [Methods (PCA+quantile space)] PCA approach: the manuscript does not state how many principal components are retained, what fraction of variance they explain, or how the reduced space is inverted back to boundary Fourier coefficients, leaving the dimensionality-reduction benefit unquantified.
minor comments (2)
  1. Notation for the quantile and PCA transforms should be defined with explicit equations rather than prose descriptions.
  2. The dataset of existing stellarator boundaries used to fit the transforms should be referenced with a table or citation so readers can assess its coverage.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and will revise the manuscript to strengthen the presentation of our results.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that the optimizations 'yield stellarator configurations with excellent fast-particle confinement' is unsupported by any quantitative metrics (loss fractions, comparison baselines, or uncertainty estimates), which is load-bearing for the central result.

    Authors: We agree the abstract would benefit from explicit quantitative support. The manuscript body reports loss fractions and comparisons for the optimized configurations; we will revise the abstract to include specific metrics such as achieved loss fractions, baseline comparisons, and uncertainty estimates drawn from the optimization results. revision: yes

  2. Referee: [Abstract] Abstract and methods description: no quantitative report is given on the fraction of proposals that produced self-intersecting surfaces or caused VMEC (or equivalent) equilibrium failures during the Bayesian search; without this, it is impossible to confirm that the quantile/PCA transforms keep the optimizer inside the valid domain.

    Authors: This is a fair point. The transforms are constructed from the existing dataset to favor valid boundaries, but the current draft lacks a reported success rate. We will add the observed fraction of invalid proposals (self-intersections or equilibrium failures) to the methods section in revision, based on the optimization logs. revision: yes

  3. Referee: [Methods (PCA+quantile space)] PCA approach: the manuscript does not state how many principal components are retained, what fraction of variance they explain, or how the reduced space is inverted back to boundary Fourier coefficients, leaving the dimensionality-reduction benefit unquantified.

    Authors: We will update the methods section to specify the number of retained principal components, the fraction of variance explained, and the explicit inversion procedure from the reduced space to Fourier coefficients. This will quantify the dimensionality reduction. revision: yes

Circularity Check

0 steps flagged

No circularity detected; parameter spaces built from external dataset via standard transforms, objective evaluated independently via guiding-center tracing.

full rationale

The paper constructs new bounded parameter spaces by applying quantile transforms (and optionally PCA) to a dataset of existing stellarator boundaries, then performs Bayesian optimization whose objective (alpha-particle losses) is computed afresh inside the loop using guiding-center tracing. No derivation step reduces to a fitted parameter being renamed as a prediction, no self-citation is load-bearing for the central result, and no uniqueness theorem or ansatz is smuggled in. The reported configurations are outputs of the optimization, not equivalent to the input dataset by construction. This is the normal case of an independent computational search.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The approach rests on standard statistical methods applied to a pre-existing dataset of stellarator boundaries. No free parameters are introduced in the abstract description, and no new physical entities are postulated.

axioms (2)
  • standard math Principal component analysis applied to sampled boundary points yields a reduced basis that preserves sufficient expressiveness for optimization.
    Invoked for the second parameter space.
  • domain assumption Quantile transformation of Fourier coefficients maps the empirical distribution to uniform [0,1] bounds while retaining optimization utility.
    Core transformation for both spaces.

pith-pipeline@v0.9.1-grok · 5792 in / 1229 out tokens · 21003 ms · 2026-06-26T18:39:48.124379+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

49 extracted references · 2 linked inside Pith

  1. [1]

    Jang B, Landreman M and Conlin R 2026Journal of Plasma Physics92E15

  2. [2]

    Landreman M and Paul E 2022Physical Review Letters128035001

  3. [3]

    Goodman A G, Mata K C, Henneberg S A, Jorge R, Landreman M, Plunk G, Smith H, Mack- enbach R, Beidler C and Helander P 2023Journal of Plasma Physics89905890504

  4. [4]

    Giuliani A 2024Journal of Plasma Physics90905900303

  5. [5]

    Dudt D W, Goodman A G, Conlin R, Panici D and Kolemen E 2024Journal of Plasma Physics 90905900120 23 Landremanet al

  6. [6]

    Liu H, Yu G, Zhu C, Velasco J L, Gaur R, Panici D, Kolemen E, Yu M, Ding W, Wang S and Zhuang G 2026 Optimizing stellarators with hidden symmetry (Preprint2502.09350) URL https://arxiv.org/abs/2502.09350

    arXiv 2026

  7. [7]

    Bader A, Anderson D, Drevlak M, Faber B, Hegna C, Henneberg S, Landreman M, Schmitt J, Suzuki Y and Ware A 2021Nuclear Fusion61116060

  8. [8]

    Paul E J, Bhattacharjee A, Landreman M, Alex D, Velasco J and Nies R 2022Nuclear Fusion 62126054

  9. [9]

    Wiedman A, Buller S and Landreman M 2024Journal of Plasma Physics90905900307

  10. [10]

    Hegna C, Anderson D, Andrew E, Ayilaran A, Bader A, Bohm T, Mata K C, Canik J, Carbajal L, Cerfon Aet al.2025Journal of Plasma Physics91E76

  11. [11]

    Landreman M, Choi J Y, Alves C, Balaprakash P, Churchill M, Conlin R and Roberg-Clark G 2025Journal of Plasma Physics91

  12. [12]

    Czekanski M, Knyazev A R, Bindel D and Paul E J 2026 CATAPULT: A CUDA-Accelerated Timestepper for Alpha Particles Using Local Tricubics (Preprint2604.07617) URLhttps: //arxiv.org/abs/2604.07617

    Pith/arXiv arXiv 2026

  13. [13]

    Bindel D, Landreman M and Padidar M 2023Plasma Physics and Controlled Fusion65065012

  14. [14]

    Velasco J, Calvo I, Escoto F, S´ anchez E, Thienpondt H and Parra F 2024Physical review letters 133185101

  15. [15]

    Giuliani A, Rodr´ ıguez E and Spivak M 2025Journal of Plasma Physics91E128

  16. [16]

    Cadena S A, Merlo A, Laude E, Bauer A, Agrawal A, Pascu M, Savtchouk M, Guiraud E, Bonauer L, Hudson Set al.2026Advances in Neural Information Processing Systems38

  17. [17]

    Wei X, Huang H, Chen H, Zhu H, Bai Z, Williams S and Lin Z 2026arXiv preprint arXiv:2603.17366

    Pith/arXiv arXiv

  18. [18]

    Ali I, Goodman A and Zingg D 2025 A B-spline parametrization for stellarator boundary opti- mization Poster presented at the 67th APS Division of Plasma Physics Meeting

    2025

  19. [19]

    Glas S, Padidar M, Kellison A and Bindel D 2022Journal of Plasma Physics88905880208

  20. [20]

    Gori S, N¨ uhrenberg J, Zille R, Okamura S, Matsuoka K and Murakami S 2001Plasma physics and controlled fusion43137–144

  21. [21]

    Ku L P, Garabedian P, Lyon J, Turnbull A, Grossman A, Mau T, Zarnstorff M and Team A 2008Fusion Science and Technology54673–693

  22. [22]

    Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, Vanderplas J, Passos A, Cournapeau D, Brucher M, Perrot M and Duchesnay E 2011Journal of Machine Learning Research122825–2830

  23. [23]

    Anderson D T and Garabedian P R 1994Nuclear Fusion34881–885

  24. [24]

    Garabedian P and Gardner H 1995Physics of Plasmas22020–2025

    2025

  25. [25]

    Dudt D and Kolemen E 2020Physics of Plasmas27

  26. [26]

    Schilling J 2025arXiv preprint arXiv:2502.04374

    arXiv

  27. [27]

    Kappel J, Landreman M and Malhotra D 2024Plasma Physics and Controlled Fusion66025018

  28. [28]

    S´ anchez E, Velasco J, Calvo I and Mulas S 2023Nuclear Fusion63066037

  29. [29]

    Hirshman S and Meier H 1985The Physics of Fluids281387–1391

  30. [30]

    Jorge R, Plunk G, Drevlak M, Landreman M, Lobsien J F, Mata K C and Helander P 2022 Journal of Plasma Physics88175880504

    2022

  31. [31]

    Alonso J, Calvo I, Carralero D, Velasco J, Garc´ ıa-Rega˜ na J, Palermo I and Rapisarda D 2022 Nuclear Fusion62036024 24 Landremanet al

    2022

  32. [32]

    Lion J, Angl` es J C, Bonauer L, Navarro A B, Ceron S C, Davies R, Drevlak M, Foppiani N, Geiger J, Goodman Aet al.2025Fusion Engineering and Design214114868

  33. [33]

    Guttenfelder W, Mandell N, Le Bars G, Singh L, Bader A, Mata K C, Canik J, Carbajal L, Cerfon A, Davila Net al.2025Journal of Plasma Physics91E83

  34. [34]

    Schmitt J, Anderson D, Andrew E, Bader A, Mata K C, Canik J, Carbajal L, Cerfon A, Cooper W, Davila Net al.2025Journal of Plasma Physics91E88

  35. [35]

    Landreman M, Torreblanca H and Cerfon A 2026Fusion Engineering and Design224115627

  36. [36]

    Swanson C, Kumar S, Dudt D, Flom E, Kalb W, Kruger T, Martin M, Olatunji J, Pasmann S, Tang Let al.2025arXiv preprint arXiv:2512.08027

    arXiv

  37. [37]

    Najmabadi F, Raffray A, Abdel-Khalik S, Bromberg L, Crosatti L, El-Guebaly L, Garabedian P, Grossman A, Henderson D, Ibrahim Aet al.2008Fusion Science and Technology54655–672

  38. [38]

    Schuett T M and Henneberg S A 2024Physical Review Research6L042052

  39. [39]

    Olson M, Santorella E, Tiao L C, Cakmak S, Garrard M, Daulton S, Lin Z J, Ament S, Beck- erman B, Onofrey E, Igusti P, Lara C, Letham B, Cardoso C, Shen S S, Lin A C, Grange M, Kashtelyan E, Eriksson D, Balandat M and Bakshy E 2025 Ax: A Platform for Adaptive ExperimentationAutoML 2025 ABCD Track

    2025

  40. [40]

    Hvarfner C, Hellsten E O and Nardi L 2024 Vanilla bayesian optimization performs great in high dimensionsProceedings of the 41st International Conference on Machine Learningpp 20793– 20817

    2024

  41. [41]

    Eriksson D, Pearce M, Gardner J, Turner R D and Poloczek M 2019Advances in neural infor- mation processing systems32

  42. [42]

    Landreman M 2026https: // doi. org/ 10. 5281/ zenodo. 20733436

  43. [43]

    Velasco J, S´ anchez E and Calvo I 2025Nuclear Fusion65056012

  44. [44]

    Liu H, Yu G, Velasco J L and Zhu C 2026arXiv preprint arXiv:2603.12139

    arXiv

  45. [45]

    Velasco J L, Calvo I, Fern´ andez-Pacheco V, Padidar M, Liu H, S´ anchez E, Yu G and Zhu C 2026arXiv preprint arXiv:2603.12377

    arXiv

  46. [46]

    Goodman A G, Xanthopoulos P, Plunk G G, Smith H, N¨ uhrenberg C, Beidler C D, Henneberg S A, Roberg-Clark G, Drevlak M and Helander P 2024PRX Energy3023010

  47. [47]

    Law F, Cerfon A and Peherstorfer B 2022Nuclear Fusion62076019

  48. [48]

    Landreman M and Jorge R 2020Journal of Plasma Physics86905860510

  49. [49]

    Panici D, Conlin R, Dudt D W, Unalmis K and Kolemen E 2023Journal of Plasma Physics89 955890303 25