Pith. sign in

REVIEW 4 major objections 5 minor 88 references

A Supervised Machine Learning Framework for Multipactor Breakdown Prediction in High-Power Radio Frequency Devices and Accelerator Components: A Case Study in Planar Geometry

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Supervised machine learning predicts multipactor susceptibility in planar RF components from voltage, frequency-gap product, and SEY parameters.

desk verdict A genuine leave-one-material-out ML benchmark for multipactor susceptibility, but every headline number is measured against spline-interpolated PIC charts, so treat the accuracy claims as suggestive, not definitive. read the letter →

arxiv 2507.17881 v1 pith:22C7J3UH submitted 2025-07-23 physics.acc-ph cs.LGphysics.app-phphysics.plasm-ph

Asif Iqbal , John Verboncoeur , Peng Zhang This is my paper
classification physics.acc-phcs.LGphysics.app-phphysics.plasm-ph
keywords multipactormachinelearningsecondaryelectronyieldparticle-in-cellsimulationsusceptibilitypredictionRandomForestsurrogatemodelingRFbreakdown
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims to present the first application of supervised machine learning to predict multipactor susceptibility, the onset of a damaging electron avalanche in high-power RF and accelerator components. Using 3D particle-in-cell simulations for six synthetic materials, the authors train regression models to map voltage, frequency-gap product, and secondary electron yield parameters to the time-averaged electron growth rate. They report that tree-based models, especially Random Forest, achieve predictive accuracy comparable to Monte Carlo simulations on a held-out copper-like material. The paper also finds that neural networks require a scalarized IoU+SSIM objective to produce structurally faithful susceptibility maps, and that feature-space separation between materials limits generalization. A sympathetic reader would accept this as a proof-of-concept that ML surrogates can replace expensive parametric PIC sweeps for planar multipactor, provided the ground-truth charts are reliable.

What carries the argument

The central object is the multipactor susceptibility chart: a two-dimensional map of the time-averaged electron growth rate $\delta_{avg}$ over the voltage ($V_{rf}$) and frequency-gap product ($f \cdot d$) plane, following Vaughan's similarity scaling. Susceptible regions are where $\delta_{avg} > 1$. The models learn a regression function $\delta_{avg} = g(V_{rf}, f\cdot d, \delta_{max}^0, E_1, E_2)$, with $E_{max}^0$ pruned after mutual information analysis. The evaluation machinery uses IoU to score spatial overlap of predicted versus PIC-simulated susceptible regions, SSIM for structural similarity of the full charts, and Pearson correlation for trend agreement. The scalarized loss $\Gamma(\theta_{mlp}) = -(0.5 \cdot IoU_{cv} + 0.5 \cdot SSIM_{cv})$ is the mechanism that makes MLPs learn physically meaningful susceptibility distributions instead of overfitting to one metric.

What would settle it

Run the trained RF model on raw, non-interpolated PIC simulation points (the original 51 by 24 grid per material) and compare IoU/SSIM against predictions on the interpolated grid; if IoU drops substantially on raw points, the reported fidelity is an artifact of smoothing. Alternatively, compare predicted susceptibility boundaries against a new experimental planar-geometry multipactor measurement for one material: any large mismatch in onset voltage or gap range would falsify the claim of MC-comparable accuracy.

Watch

Extended reading notes

Core claim

The central claim is that a supervised machine learning framework can learn the mapping from operational and material features to multipactor susceptibility in a two-surface planar geometry, producing predictions accurate enough for design screening. The authors establish this by training Random Forest, Extra Trees, XGBoost, and funnel-structured MLPs on a dataset derived from 3D PIC simulations, with leave-one-material-out cross-validation to test generalization. Key results include: RF and ET outperform MLPs across IoU, SSIM, and Pearson metrics; MLPs optimized with a scalarized loss combining IoU and SSIM outperform those using either metric alone; and removing the SEY parameter $E_max^{0}$, justified by low mutual information, does not degrade performance. Benchmarking on a copper material, the RF model achieves IoU=0.61, SSIM=0.83, and Pearson=0.90, compared to MC-PIC ground-truth scores of IoU=0.65, SSIM=0.94, and Pearson=0.95, which the authors interpret as comparable predictive accuracy at far lower cost.

Load-bearing premise

The evaluation treats the spline-interpolated, space-charge-free 3D PIC susceptibility charts as the ground truth, so if that interpolation smooths away fine multipactor bands or if the PIC model misses physics, every reported model accuracy inherits that error.

Editorial extensions

If this is right

  • If the central claim is correct, design engineers can replace expensive PIC scans over voltage and gap distance with trained ML surrogates that return predicted susceptibility maps in milliseconds, enabling rapid screening of operating windows.
  • The demonstrated RF accuracy on a held-out material suggests that, with a larger dataset of SEY profiles, tree-based surrogates could cover realistic material libraries for planar geometries and potentially extend to stripline and coaxial geometries.
  • The finding that $E_{max}^0$ is redundant when crossover energies are included implies that future multipactor datasets need only report $\delta_{max}^0$, $E_1$, and $E_2$, simplifying both simulation and experimental characterization efforts.
  • The success of the scalarized IoU+SSIM objective for MLPs indicates that future neural-network surrogates for multipactor should be trained on region-overlap and structural metrics rather than pure regression loss, a design choice that directly transfers to other physics susceptibility problems.
  • The PCA-based diagnosis of performance drops for materials M4 and M5 provides a concrete criterion for dataset construction: materials should be chosen to fill the SEY feature space uniformly, avoiding isolated clusters that force extrapolation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's benchmark against MC simulations is closer than it appears: the MC-PIC IoU of 0.65 is the baseline agreement between two different simulation methodologies, not an absolute truth, so the RF's IoU of 0.61 indicates the ML model captures as much of the PIC chart as the MC simulation does.
  • The interpolation of the 51x24 PIC grid to 10,000 points per chart likely smooths fine susceptibility bands; a direct comparison of predictions against the raw 1,224 simulated points (or a held-out subset of them) would test whether the reported IoU scores overstate fidelity to actual PIC results.
  • The leave-one-material-out protocol is a stringent extrapolation test, but the paper's own PCA suggests why industry adoption may require a different validation: in practice, a new device's material is often known from the literature, so interpolating between nearby material profiles (rather than extrapolating to a cluster) would be the more common and easier use case.
  • A testable extension is to quantify gains over analytical Vaughan threshold curves: since Vaughan's similarity scaling already provides first-order susceptibility boundaries, a direct comparison of ML IoU against Vaughan-predicted boundaries on the same charts would establish the added value of the data-driven approach.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes a supervised machine learning framework for predicting multipactor susceptibility in a planar two-surface geometry. A dataset is generated with 3D PIC simulations (CST Particle Studio) for six synthetic materials with distinct SEY parameters, each on a 51x24 grid over (V_rf, fd), then spline-interpolated to 10,000 points per chart. Models (Random Forest, Extra Trees, XGBoost, and funnel-structured MLPs with Bayesian hyperparameter optimization) are trained to regress the time-averaged electron growth rate δ_avg from features (V_rf, fd, δ_max, E1, E2), under a leave-one-material-out cross-validation scheme. Performance is evaluated with IoU, SSIM, and Pearson correlation. The authors report that tree ensembles outperform MLPs, that a scalarized IoU+SSIM objective improves MLP behavior, and that the RF model achieves accuracy comparable to a published 2D Monte Carlo simulation (IoU 0.61 vs 0.65 on material M2). PCA is used to explain poor generalization on materials M4 and M5 via disjoint feature-space clusters. The central claims are that ML can serve as a fast surrogate for PIC-based susceptibility sweeps and that this is the first ML application to multipactor susceptibility prediction.

Significance. If the reported accuracy holds against the raw simulation outputs, the framework is a useful proof of concept for fast multipactor susceptibility screening, which is currently computationally expensive. The leave-one-material-out evaluation is a genuine out-of-distribution test and is more rigorous than a random train/test split. The paper is also transparent about known limitations (small dataset, extrapolation difficulty, feature-space isolation) and provides a useful comparison of model families. However, the significance is materially contingent on the fact that all quantitative claims are evaluated against spline-interpolated versions of the PIC charts rather than raw PIC outputs or experimental measurements; the 'comparable to MC' benchmark is a comparison of two agreements with the same interpolated reference. The contribution is therefore a demonstration of ML surrogate modeling for a smooth interpolant of a physics simulation output, and the paper would be strengthened by validating against the raw 51x24 grid and providing uncertainty quantification.

major comments (4)
  1. [§2.1, §2.8, Fig. 1 and Figs. 4–13] All reported IoU, SSIM, and Pearson values are computed on the 10,000-point nonlinear spline interpolations of the 51x24 PIC grids, not on the raw simulation outputs. Since the regression target is the interpolant, a model can score well by learning the smooth spline structure while missing fine susceptibility bands. The authors should report the same metrics evaluated on the original 1224 PIC sample points per material (or at least a stratified subset) and quantify the interpolation error, e.g., RMSE between the interpolated and raw δ_avg values at the grid points. Without this, the headline numbers only characterize agreement with an interpolation surrogate.
  2. [§3.1] The claim that the ML model achieves accuracy 'comparable to computationally intensive MC simulations' is based on comparing RF IoU=0.61 with a 2D MC IoU=0.65, where both quantities are agreements with the same interpolated 3D PIC chart for material M2. The MC reference was digitized from a published figure, with no documented uncertainty from digitization, and neither the RF nor the MC value carries error bars or sensitivity estimates (e.g., to interpolation density, PIC duration, or MC seed). The benchmark should be rephrased with these caveats, and ideally the RF prediction should be compared directly to the raw PIC chart points used to build the M2 reference.
  3. [§2.5 and §3.1, Fig. 6] The decision to prune E_max was made after computing mutual information on the full six-material dataset, which includes the material that is subsequently held out in each leave-one-material-out fold. This leaks test-material information into the feature-selection step and contradicts the statement that the held-out chart is excluded from all stages of model selection. The redundancy of E_max should be verified by recomputing MI on the five training materials only within each fold, or by reframing the claim as a post-hoc observation rather than a cross-validated feature-selection result.
  4. [§2.1 and abstract] The PIC simulations explicitly turn off space-charge effects, while the cited prior work (Iqbal et al., 2022) includes space-charge saturation. Multipactor susceptibility boundaries in two-surface geometries are known to be affected by space charge, especially at high growth rates. The surrogate is therefore only validated for the space-charge-free model, and the abstract's broad claim about predicting multipactor susceptibility in high-power RF devices is not fully supported. Adding a comparison with a space-charge-included PIC run for at least one material would clarify the regime of validity.
minor comments (5)
  1. [§3.3.3, Fig. 12] The text says 'panels 11(c), (e), and (f)' when referring to the training/validation loss curves; these panels belong to Fig. 12, not Fig. 11.
  2. [§2.6.3, Eq. (4)] Eq. (4) contains a stray closing parenthesis after the bias term; the equation should read X^(l)=σ(W^(l)X^(l−1)+b^(l)).
  3. [§2.5, Eq. (3)] The mutual information formula in Eq. (3) is garbled in the typesetting (the integrand appears as 'logJp(x,y)p(x)p(y)K'); it should be rewritten with proper logarithmic notation.
  4. [Table 3 and §2.6.3] Table 3 reports the learning-rate range with corrupted superscripts ('10^>' and '10^)'); the text in §2.6.3 states the intended range of 10^-3 to 10^-1. Also, the 50-epoch and batch-size 256 settings used in training are not listed in Table 3.
  5. [§1, introduction] The claim 'this work, to our knowledge, presents the first application of machine learning for predicting multipactor susceptibility' is made without a targeted literature search; the cited ML-accelerator references are broad. The authors should temper the claim with a specific search strategy or acknowledge any prior surrogate-modeling work for vacuum breakdown.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core leave-one-material-out evaluation is a genuine holdout, and the cited prior-work MC benchmark is independent and not load-bearing.

full rationale

The paper's central derivation is an empirical supervised-learning pipeline, not a closed-form derivation whose output is equivalent to its input by construction. The dataset (Section 2.1) is generated from 3D PIC simulations and then spline-interpolated to 10,000 points per chart. The models are trained to map (V, fd, delta_max, E_max, E1, E2) to the time-averaged growth rate delta_avg, and performance is scored against the held-out material's interpolated PIC chart. Section 2.7 explicitly states that the held-out chart is 'excluded from all stages of hyperparameter tuning,' so the main IoU/SSIM/Pearson results are genuine out-of-sample predictions rather than fitted-input predictions. The spline interpolation is a target-generation choice, not a mechanism that makes the model's output equal to its input; the same interpolated surface is used for all models, so any interpolation artifact affects all methods equally and does not force the claimed rankings. The MC benchmark in Section 3.1 uses a published chart from Iqbal et al. (2022), a self-citation by the same authors, but the MC simulation is an independent prior computation with stated assumptions, and the comparison serves as a secondary anchor rather than as the source of the model predictions. The E_max pruning decision based on mutual information computed over all six charts is a feature-selection-on-test-data concern that could bias reported generalization, but it is not a circular reduction: the target is not redefined in terms of the model, and no fitted parameter is renamed as a prediction. The paper also openly acknowledges its own limitations (Section 3.3.4, small dataset and out-of-distribution challenges), further supporting the assessment that the central claims are not smuggled in through self-citation or by definition. No circular step meeting the required evidentiary standard was found.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The paper contributes a data-generation and benchmarking exercise, so it leans heavily on prior physics models (Vaughan SEY, PIC method) and on hand-chosen simulation settings. No new physical entity is introduced. The main ledger entries are the modeling assumptions that define the ground truth, since every ML accuracy number is measured against those charts.

free parameters (4)
  • MLP hyperparameters (number of layers, initial neurons, activation, L2 alpha, learning rate) = Final per-material values not reported
    Selected by Bayesian optimization during 5-fold cross-validation; the MLP comparisons depend on these choices.
  • Tree ensemble hyperparameters (RF, ET, XGB settings) = Not reported; likely library defaults
    The claimed superiority of tree-based models could be affected by unstated hyperparameter choices.
  • Six synthetic material SEY profiles (M1-M6) = Values in Table 1
    Chosen by hand to span SEY space; generalization claims rest on this small, clustered selection.
  • Spline interpolation density (10,000 points per chart) = 10,000 points per chart
    Chosen as augmentation level; can introduce smoothness artifacts that become part of the evaluation target.
assumptions (6)
  • domain assumption Vaughan SEY model accurately represents secondary emission for the six materials
    Used in Section 2.1 to define material response and generate PIC data; if inaccurate, susceptibility maps shift.
  • domain assumption Multipactor susceptibility depends only on V, fd, and SEY parameters as in Eq. 2
    Inherits Vaughan's planar similarity scaling assumptions; reduces geometry to two operational inputs.
  • domain assumption 3D PIC simulations with space charge off and 10 ns runtime give converged, accurate ground truth
    Invoked in Section 2.1; no experimental validation or convergence study is provided.
  • domain assumption Spline interpolation within the simulated domain preserves PIC fidelity without physical artifacts
    Section 2.1 uses it to create 10,000 points per chart; all evaluation metrics are computed against this interpolated target.
  • domain assumption Leave-one-material-out with six clusters approximates real-world material generalization
    Section 2.7 treats each material chart as one group; with only six materials, each held-out test is effectively one cluster point.
  • standard math Standard ML and statistical tools (MI, PCA, cross-validation, tree ensembles) are valid for this task
    Sections 2.5 to 2.8 rely on established libraries and methods with known assumptions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Supervised Machine Learning Framework for Multipactor Breakdown Prediction in High-Power Radio Frequency Devices and Accelerator Components: A Case Study in Planar Geometry." pith.science (2026). https://pith.science/paper/22C7J3UH

@misc{pith2026250717881,
  author       = {Pith},
  title        = {Pith review of: A Supervised Machine Learning Framework for Multipactor Breakdown Prediction in High-Power Radio Frequency Devices and Accelerator Components: A Case Study in Planar Geometry},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/22C7J3UH}},
  note         = {Machine review of arXiv:2507.17881}
}
abstract

Multipactor is a nonlinear electron avalanche phenomenon that can severely impair the performance of high-power radio frequency (RF) devices and accelerator systems. Accurate prediction of multipactor susceptibility across different materials and operational regimes remains a critical yet computationally intensive challenge in accelerator component design and RF engineering. This study presents the first application of supervised machine learning (ML) for predicting multipactor susceptibility in two-surface planar geometries. A simulation-derived dataset spanning six distinct secondary electron yield (SEY) material profiles is used to train regression models - including Random Forest (RF), Extra Trees (ET), Extreme Gradient Boosting (XGBoost), and funnel-structured Multilayer Perceptrons (MLPs) - to predict the time-averaged electron growth rate, ${\delta}_{avg}$. Performance is evaluated using Intersection over Union (IoU), Structural Similarity Index (SSIM), and Pearson correlation coefficient. Tree-based models consistently outperform MLPs in generalizing across disjoint material domains. MLPs trained using a scalarized objective function that combines IoU and SSIM during Bayesian hyperparameter optimization with 5-fold cross-validation outperform those trained with single-objective loss functions. Principal Component Analysis reveals that performance degradation for certain materials stems from disjoint feature-space distributions, underscoring the need for broader dataset coverage. This study demonstrates both the promise and limitations of ML-based multipactor prediction and lays the groundwork for accelerated, data-driven modeling in advanced RF and accelerator system design.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

88 extracted references · 66 canonical work pages

  1. [1]

    Introduction Multipactor discharge (Iqbal et al., 2018; Kishek et al., 1998; Vaughan,

    2018

  2. [2]

    The pipeline begins with PIC simulations in CST Particle Studio to generate a multipactor susceptibility dataset for six materials

    Workflow of the supervised machine learning framework for multipactor susceptibility prediction. The pipeline begins with PIC simulations in CST Particle Studio to generate a multipactor susceptibility dataset for six materials. Input features include operational and geometric variables (V$%, f, d), and material-dependent secondary electron yield (SEY) pr...

    work page 1988

  3. [3]

    #. Scores are normalized by the total mutual information across all features such that the bars sum to 1, highlighting each feature’s relative importance in predicting δ!

    Normalized mutual information (MI) scores indicating the relative contribution of each input feature to predicting the electron growth rate δ!"#. Scores are normalized by the total mutual information across all features such that the bars sum to 1, highlighting each feature’s relative importance in predicting δ!"#. Fig. 3 presents the relative mutual info...

    work page 2012

  4. [6]

    that described basic mechanisms of secondary electron avalanches. However, these models often made simplifying assumptions about geometry, field structure, and emission dynamics, limiting their applicability to practical devices with complex shapes and variable operating conditions. With advances in computational capabilities, more sophisticated technique...

    2020

  5. [9]

    # in the (V$%, f×d) plane (Fig. 1), following the similarity scaling law of Vaughan (1988), where f is the excitation frequency. Regions where δ!

    as implemented in CST PS, with emission energy and angular distributions of secondary electrons following the approach described by Iqbal et al. (2023b, 2022). Space-charge effects were turned off, allowing the investigation of pure secondary electron-driven multipactor growth without space-charge suppression (Iqbal et al., 2023b, 2022). For each combinat...

    work page 2022

  6. [12]

    Entropy, Relative Entropy, and Mutual Information,

    as described in Section 2.1. Consequently, the susceptibility parameter is modeled as: δ!"#=𝑔(V$%,fd,δ&!'(,E&!'(,E),E*) (2) The selection of these input features is motivated by a range of previous experimental, analytical, and simulation studies that established their critical roles in determining multipactor onset and growth (Iqbal et al., 2023a; Kishek...

    work page 1998

  7. [16]

    # is shown in the color bar. Regions with δ!

    Results and Discussion This section presents the results of machine learning model performance in predicting multipactor susceptibility. We begin by evaluating RF model predictions and benchmarking against Monte Carlo simulations. We also evaluate the impact of including E&!'( as an input feature. Subsequently, we present susceptibility prediction results...

    work page 2022

  8. [17]

    Under the leave-one-material-out cross-validation strategy, the held-out material occupies a region in feature space that is entirely disjoint from the training data

    The projection reveals that susceptibility data points from different materials form distinct, non-overlapping clusters in the reduced PCA space due to variations in material dependent SEY features. Under the leave-one-material-out cross-validation strategy, the held-out material occupies a region in feature space that is entirely disjoint from the traini...

    work page 2020

Show all 88 references
  1. [18]

    In contrast, ensemble tree models like Random Forests, due to their localized decision structures, often exhibit more stable performance in heterogeneous or sparsely sampled domains (Fernández-Delgado et al., 2014; Zhou and Feng, 2017). 27 Fig

    2014

  2. [23]

    Bagging predictors. Mach. Learn. 24, 123–140. https://doi.org/10.1007/BF00058655 Bruning, O., Caspers, F., Collins, I.R., Grobner, O., Henrist, B., Hilleret, N., Laurent, J.-M., Morvillo, M., Pivi, M., Ruggiero, F., Zhang, X.,

    doi

  3. [24]

    No.99CH36366)

    Electron cloud and beam scrubbing in the LHC, in: Proceedings of the 1999 Particle Accelerator Conference (Cat. No.99CH36366). Presented at the Proceedings of the 1999 Particle Accelerator Conference (Cat. No.99CH36366), IEEE, New York, NY, USA, pp. 2629–2631 vol.4. https://do...

    1999

  4. [31]

    Plasma cleaning of the SLAC Linac Coherent Light Source II high energy verification cryomodule cavities. Phys. Rev. Accel. Beams 25, 102001. https://doi.org/10.1103/PhysRevAccelBeams.25.102001 Gill, E.W.B., von Engel, A.,

    doi

  5. [33]

    https://doi.org/10.48550/arXiv.2007.01434 Gutton, C., Gutton, H.,

    In Search of Lost Domain Generalization. https://doi.org/10.48550/arXiv.2007.01434 Gutton, C., Gutton, H.,

    2007 doi arXiv

  6. [36]

    Empirical modeling and Monte Carlo simulation of secondary electron yield reduction of laser drilled microporous gold surfaces. J. Vac. Sci. Technol. B 38, 013801. https://doi.org/10.1116/1.5130683 Iqbal, A., Verboncoeur, J., Zhang, P.,

    doi

  7. [37]

    Two surface multipactor discharge with two-frequency rf fields and space-charge effects. Phys. Plasmas 29, 012102. https://doi.org/10.1063/5.0070889 Iqbal, A., Verboncoeur, J., Zhang, P.,

    doi

  8. [38]

    Multipactor susceptibility on a dielectric with two carrier frequencies. Phys. Plasmas 25, 043501. https://doi.org/10.1063/1.5024365 Iqbal, A., Wen, D., Verboncoeur, J., Zhang, P., 2023a. Recent advances in multipactor physics and mitigation. High Volt. 8, 1095–1114. https://d...

    doi

  9. [39]

    Time-dependent physics of single-surface multipactor discharge with two carrier frequencies. Phys. Rev. E 102, 043201. https://doi.org/10.1103/PhysRevE.102.043201 Iqbal, Asif,

    doi

  10. [40]

    Michigan State University

    MULTIPACTOR DISCHARGE WITH TWO-FREQUENCY RF FIELDS. Michigan State University. https://doi.org/10.25335/K2V7-YA15 Jeong, Y., Jo, J., Lee, T., Yoo, J.,

    doi

  11. [41]

    Combined analysis of thermofluids and electromagnetism using physics-informed neural networks. Eng. Appl. Artif. Intell. 133, 108216. https://doi.org/10.1016/j.engappai.2024.108216 Jing, C., Chang, C., Gold, S.H., Konecny, R., Antipov, S., Schoessow, P., Kanareykin, A., Gai, W.,

    2024

  12. [43]

    Cheetah: Bridging the Gap Between Machine Learning and Particle Accelerator Physics with High-Speed, Differentiable Simulations. Phys. Rev. Accel. Beams 27, 054601. https://doi.org/10.1103/PhysRevAccelBeams.27.054601 Kang,Yoon, Kim,Sang-Ho, Lee,Sung-Woo, Ren,Haitao, Moss,John,...

    doi

  13. [48]

    Multipactor discharge on metals and dielectrics: Historical review and recent theories. Phys. Plasmas 5, 2120–2126. https://doi.org/10.1063/1.872883 Kohavi, R.,

    doi

  14. [49]

    Multipactor experiments on an S-band coaxial test cell. Rev. Sci. Instrum. 92, 124706. https://doi.org/10.1063/5.0074464 Li, Y., Lei, G., Bramerdorfer, G., Peng, S., Sun, X., Zhu, J.,

    doi

  15. [51]

    Quantitative analysis of multipactor threshold sensitivity to secondary emission yield of microwave devices. Phys. Plasmas 30, 033104. https://doi.org/10.1063/5.0138875 Ludwick, J., Iqbal, A., Gortat, D., Cook, J.D., Cahay, M., Zhang, P., Back, T.C., Fairchild, S., Sparkes, M....

    doi

  16. [52]

    Angular dependence of secondary electron yield from microporous gold surfaces. J. Vac. Sci. Technol. B 38, 054001. https://doi.org/10.1116/6.0000346 Macek, R., Wei, J., Hseuh, H.C., Blaskiewicz, M., Zhang, S.Y.,

    doi

  17. [54]

    Ultrawideband, high-power, microstripline test setup for experimental study and characterization of multipactor. Rev. Sci. Instrum. 92, 084706. https://doi.org/10.1063/5.0058049 Mirmozafari, M., Iqbal, A., Zhang, P., Behdad, N., Booske, J.H., Verboncoeur, J.P.,

    doi

  18. [55]

    Secondary electron yield characterization of high porosity surfaces for multipactor-free microwave components. Phys. Plasmas 29, 082109. https://doi.org/10.1063/5.0095160 Moss, J., Crofford, M., Lee, S.-W., Middendorf, M., Toby, G.,

    doi

  19. [56]

    The Spallation Neutron Source Normal Conducting Linac RF System Design for the Proton 33 Power Upgrade Project. Proc. 12th Int. Part. Accel. Conf. IPAC2021, 4 pages, 0.722 MB. https://doi.org/10.18429/JACOW-IPAC2021-THPAB296 Pathak, A.,

    doi

  20. [57]

    Linac_Gen: integrating machine learning and particle-in-cell methods for enhanced beam dynamics at Fermilab. https://doi.org/10.48550/arXiv.2406.16630 Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dub...

    doi arXiv

  21. [58]

    Demonstration of Niobium Tin in 218 MHz Low-beta Quarter Wave Accelerator Cavity. https://doi.org/10.48550/arXiv.2307.08783 Petersen,Troy, Guilfoyle,Berardino, Tennis,Brad, Spranza,Edward, Eremeev,Grigory, Kelly,Michael, Agustsson,Ronald, Posen,Sam, Kutsaev,Sergey, Reid,Thomas,

    doi arXiv

  22. [59]

    Niobium-tin as a transformative technology for low-beta linacs 2709-2712 pages, 0.71 MB. https://doi.org/10.18429/JACOW-IPAC2024-WEPS09 Posen, S., Cravatta, A., Checchin, M., Aderhold, S., Adolphsen, C., Arkan, T., Bafia, D., Benwell, A., Bice, D., Chase, B., Contreras-Martine...

    doi

  23. [60]

    High gradient performance and quench behavior of a verification cryomodule for a high energy continuous wave linear accelerator. Phys. Rev. Accel. Beams 25, 042001. https://doi.org/10.1103/PhysRevAccelBeams.25.042001 Power, J.G., Gold, S.H.,

    doi

  24. [61]

    AIP Conf

    Multipactor Modeling in Cylindrical Dielectric-Loaded Accelerators. AIP Conf. Proc. 877, 362–369. https://doi.org/10.1063/1.2409157 Preist, D.H., Talcott, R.C.,

    doi

  25. [62]

    IRE Trans

    On the heating of output windows of microwave tubes by electron bombardment. IRE Trans. Electron Devices 8, 243–251. https://doi.org/10.1109/T-ED.1961.14797 Romanov, G.,

    1961

  26. [64]

    Contribution to modeling multipactor and corona discharges in high power electromagnetic fields. EPFL. https://doi.org/10.5075/epfl-thesis-5542 Sagar, M.S.I., Ouassal, H., Omi, A.I., Wisniewska, A., Jalajamony, H.M., Fernandez, R.E., Sekhar, P.K.,

    doi

  27. [66]

    Taking the Human Out of the Loop: A Review of Bayesian Optimization. Proc. IEEE 104, 148–175. https://doi.org/10.1109/JPROC.2015.2494218 Shorten, C., Khoshgoftaar, T.M.,

    2015

  28. [67]

    https://doi.org/10.1186/s40537-019-0197-0 Simakov,Evgenya, Alexander,Anna, Haynes,Brian, Rai,Deepak, Xu,Haoran, Zuboraj,Muhammed, Anisimov,Petr, Grumstrup,Torben, Barkley,Walter,

    doi

  29. [68]

    https://doi.org/10.18429/JACOW-IPAC2024-WEPC60 Space Engineering: Multipacting Design and Test,

    Status of the CARIE high gradient photocathode test facility at LANL 2101-2104 pages, 2.3 MB. https://doi.org/10.18429/JACOW-IPAC2024-WEPC60 Space Engineering: Multipacting Design and Test,

    doi

  30. [69]

    Plasma Phys

    Contrib. Plasma Phys. 63, e202300060. https://doi.org/10.1002/ctpp.202300060 Spirkin, A., Stoltz, P.H., Luginsland, J.W.,

    doi

  31. [70]

    IEEE Trans

    Machine Learning Investigation of the Rising Sun Magnetron Design and Operation. IEEE Trans. Plasma Sci. 48, 133–140. https://doi.org/10.1109/TPS.2019.2958289 Tizakast, Y., Kaddiri, M., Lamsaadi, M., Makayssi, T.,

    2019

  32. [71]

    Machine Learning based algorithms for modeling natural convection fluid flow and heat and mass transfer in rectangular cavities filled with non-Newtonian fluids. Eng. Appl. Artif. Intell. 119, 105750. https://doi.org/10.1016/j.engappai.2022.105750 Toby, G., Kang, Y., Kim, S.-H...

  33. [72]

    JACoW IPAC2023, WEPA127

    Multipacting analysis of the SNS drift tube linac (DTL) RF vacuum window using Spark3d. JACoW IPAC2023, WEPA127. https://doi.org/10.18429/JACoW-IPAC2023-WEPA127 Vahedi, V., Birdsall, C.K., Lieberman, M.A., DiPeso, G., Rognlien, T.D.,

    doi

  34. [73]

    Verification of frequency scaling laws for capacitive radio-frequency discharges using two-dimensional simulations*. Phys. Fluids B Plasma Phys. 5, 2719–2729. https://doi.org/10.1063/1.860711 Vahedi, V., DiPeso, G.,

    doi

  35. [75]

    IEEE Trans

    A new formula for secondary emission yield. IEEE Trans. Electron Devices 36, 1963–1967. https://doi.org/10.1109/16.34278 Vaughan, J.R.M.,

    1963 doi

  36. [76]

    IEEE Trans

    Multipactor. IEEE Trans. Electron Devices 35, 1172–1180. https://doi.org/10.1109/16.3387 Vaughan, J.R.M.,

    doi

  37. [77]

    IRE Trans

    Some high-power window failures. IRE Trans. Electron Devices 8, 302–308. https://doi.org/10.1109/T-ED.1961.14804 Vaughan, R.M.,

    1961

  38. [78]

    IEEE Trans

    Secondary emission formulas. IEEE Trans. Electron Devices 40, 830-. https://doi.org/10.1109/16.202798 Vdovicheva, N.K., Sazontov, A.G., Semenov, V.E.,

    doi

  39. [80]

    Simultaneous Potential and Circuit Solution for 1D Bounded Plasma Particle Simulation Codes. J. Comput. Phys. 104, 321–328. https://doi.org/10.1006/jcph.1993.1034 Wong, P.Y., Lau, Y.Y., Zhang, P., Jordan, N., Gilgenbach, R.M., Verboncoeur, J.,

    1993

  40. [81]

    The effects of multipactor on the quality of a complex signal propagating in a transmission line. Phys. Plasmas 26, 112114. https://doi.org/10.1063/1.5125408 Woo, R., 1968a. Multipacting Discharges between Coaxial Electrodes. J. Appl. Phys. 39, 1528–1533. https://doi.org/10.10...

  41. [82]

    Advances in Multiphysics Modeling for Parallel Finite-Element Code Suite ACE3P. IEEE J. Multiscale Multiphysics Comput. Tech. 4, 298–306. https://doi.org/10.1109/JMMCT.2019.2954946 Xu,Haoran, Alexander,Anna, Simakov,Evgenya, Bustos,Gerald, Lawler,Gerard, Rosenzweig,James, Anis...

    2019

  42. [83]

    https://doi.org/10.18429/JACOW-IPAC2024-THPG04 Yang, S., Nam, J., Dietschreit, J.C.B., Gómez-Bombarelli, R.,

    RF and multipactor analysis for the CARIE RF photoinjector with a photocathode insert 3251-3253 pages, 4.9 MB. https://doi.org/10.18429/JACOW-IPAC2024-THPG04 Yang, S., Nam, J., Dietschreit, J.C.B., Gómez-Bombarelli, R.,

    doi

  43. [84]

    Learning Collective Variables with Synthetic Data Augmentation through Physics-Inspired Geodesic Interpolation. J. Chem. Theory Comput. 20, 6559–6568. https://doi.org/10.1021/acs.jctc.4c00435 Zhang, P., Lau, Y.Y., Franzi, M., Gilgenbach, R.M.,

    doi

  44. [86]

    Deep learning for thermal plasma simulation: Solving 1-D arc model as an example. Comput. Phys. Commun. 257, 107496. https://doi.org/10.1016/j.cpc.2020.107496 Zhou, Z.-H., Feng, J.,

    2020

  45. [87]

    Deep Forest: Towards An Alternative to Deep Neural Networks, in: Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence. Presented at the Twenty-Sixth International Joint Conference on Artificial Intelligence, International Joint Conferences ...

    doi

  46. [88]

    https://doi.org/10.48550/arXiv.2302.02092 Zou, H., Zeng, S., Li, C., Ji, J.,

    Interpolation for Robust Learning: Data Augmentation on Wasserstein Geodesics. https://doi.org/10.48550/arXiv.2302.02092 Zou, H., Zeng, S., Li, C., Ji, J.,

    doi arXiv

  47. [89]

    A survey of machine learning and evolutionary computation for antenna modeling and optimization: Methods and challenges. Eng. Appl. Artif. Intell. 138, 109381. https://doi.org/10.1016/j.engappai.2024.109381

    2024

  48. [99]

    https://doi.org/10.11648/j.ajpa.20170506.15 Rosario, S., Edén, F.,

    doi

  49. [1627]

    https://doi.org/10.3390/app11041627 Lin, S., Qu, H., Xia, N., Wong, P., Zhang, P., Verboncoeur, J., Cao, M., Zhai, Y., Li, Y., Wang, H.,

    doi

  50. [1954]

    The Secondary Electron Resonance Mechanism of Low-Pressure High-Frequency Gas Breakdown. J. Appl. Phys. 25, 417–423. https://doi.org/10.1063/1.1721656 Iqbal, A., Ludwick, J., Fairchild, S., Cahay, M., Gortat, D., Sparkes, M., O’Neill, W., Back, T.C., Zhang, P.,

    doi

  51. [1958]

    Multipacting Modes of High-Frequency Gaseous Breakdown. Phys. Rev. 112, 681–685. https://doi.org/10.1103/PhysRev.112.681 Hatch, A.J., Williams, H.B.,

    doi

  52. [1961]

    and analytical models (Gill and von Engel, 1948; Gutton and Gutton, 1928; Hatch and Williams, 1958, 1954; Vaughan,

    1948

  53. [1988]

    This highly nonlinear phenomenon is triggered by secondary electron emission (SEE) (Furman and Pivi, 2002; Iqbal et al., 2019; Ludwick et al., 2020; Vaughan, 1989,

    is a vacuum electron avalanche that occurs in high-power radio frequency (RF) and microwave (HPM) devices (Benford et al., 2025). This highly nonlinear phenomenon is triggered by secondary electron emission (SEE) (Furman and Pivi, 2002; Iqbal et al., 2019; Ludwick et al., 2020...

    2025

  54. [1989]

    5 Table 1: Secondary electron yield (SEY) parameters for the six material property sets used in this study

    to represent a range of hypothetical surfaces encountered in high-power RF and accelerator components (Table 1). 5 Table 1: Secondary electron yield (SEY) parameters for the six material property sets used in this study. Material 𝛅𝐦𝐚𝐱𝟎 𝐄𝐦𝐚𝐱𝟎 (eV) 𝐄𝟏 (eV) 𝐄𝟐 (eV) M1 2.09 165.0 ...

    1900

  55. [1993]

    and can develop under a range of conditions, including resonant interactions between electrons and the RF electric field in two-surface geometries (e.g., planar (Iqbal et al., 2022; Vaughan, 1988), stripline (Mirmozafari et al., 2022, 2021)), and coaxial (Langellotti et al., 2...

    2022

  56. [1996]

    models were employed: • Random Forest (RF) regression (Breiman, 2001): A bootstrap-aggregated ensemble of decision trees, trained with random feature selection at each split to reduce overfitting and variance. • Extra Trees (ET) regression (Geurts et al., 2006): An ensemble si...

    2001

  57. [1997]

    Simultaneous Potential and Circuit Solution for Two-Dimensional Bounded Plasma Simulation Codes. J. Comput. Phys. 131, 149–163. https://doi.org/10.1006/jcph.1996.5591 Vaughan, J.R.M.,

    1996

  58. [1998]

    Multipactor Discharge on a Dielectric. Phys. Rev. Lett. 80, 193–196. https://doi.org/10.1103/PhysRevLett.80.193 Kishek, R.A., Lau, Y.Y., Ang, L.K., Valfells, A., Gilgenbach, R.M.,

    doi

  59. [1999]

    XGBoost is a scalable, regularized boosting algorithm that sequentially fits new decision trees to correct the residual errors of previous trees

    ensemble model—Extreme Gradient Boosting (XGBoost) (Chen and Guestrin, 2016)—was also explored. XGBoost is a scalable, regularized boosting algorithm that sequentially fits new decision trees to correct the residual errors of previous trees. 2.6.3 Neural Network: Multilayer Pe...

    2016

  60. [2001]

    Random Forests. Mach. Learn. 45, 5–32. https://doi.org/10.1023/A:1010933404324 Breiman, L.,

    doi

  61. [2002]

    Probabilistic model for the simulation of secondary electron emission. Phys. Rev. Spec. Top. - Accel. Beams 5, 124404. https://doi.org/10.1103/PhysRevSTAB.5.124404 Geurts, P., Ernst, D., Wehenkel, L.,

    doi

  62. [2003]

    Presented at the Proceedings of the 2003 Particle Accelerator Conference, IEEE, Portland, OR, USA, pp

    Beam scrubbing strategy for electron-cloud suppression in the Spallation neutron source ring, in: Proceedings of the 2003 Particle Accelerator Conference. Presented at the Proceedings of the 2003 Particle Accelerator Conference, IEEE, Portland, OR, USA, pp. 3419–3421. https://...

    2003 arXiv

  63. [2004]

    Radiophys

    Statistical Theory of Two-Sided Multipactor. Radiophys. Quantum Electron. 47, 580–596. https://doi.org/10.1023/B:RAQE.0000049556.18329.e9 35 Verboncoeur, J.P., Alves, M.V., Vahedi, V., Birdsall, C.K.,

  64. [2005]

    John Wiley & Sons, Ltd, pp

    , in: Elements of Information Theory. John Wiley & Sons, Ltd, pp. 13–55. https://doi.org/10.1002/047174882X.ch2 Farnsworth, P.T.,

    doi

  65. [2006]

    Extremely randomized trees. Mach. Learn. 63, 3–42. https://doi.org/10.1007/s10994-006-6226-1 Giaccone, B., Berrutti, P., Martinello, M., Posen, S., Cravatta, A., Netepenko, A., Arkan, T., Grassellino, A., Hartsell, B., Kaluzny, J., Penhollow, A., Gonnella, D., Ross, M., Fuerst...

    doi

  66. [2010]

    Nonstationary statistical theory for multipactor. Phys. Plasmas 17, 062110. https://doi.org/10.1063/1.3443128 Anza, S., Vicente, C., Gil, J., Mattes, M., Wolk, D., Wochner, U., Boria, V.E., Gimeno, B., Raboso, D.,

    doi

  67. [2011]

    Multipactor susceptibility on a dielectric with a bias dc electric field and a background gas. Phys. Plasmas 18, 053508. https://doi.org/10.1063/1.3592990 Zhong, L., Gu, Q., Wu, B.,

    doi

  68. [2012]

    IEEE Trans

    Prediction of Multipactor Breakdown for Multicarrier Applications: The Quasi-Stationary Method. IEEE Trans. Microw. Theory Tech. 60, 2093–2105. https://doi.org/10.1109/TMTT.2012.2197021 Arjovsky, M., Bottou, L., Gulrajani, I., Lopez-Paz, D.,

    2012

  69. [2013]

    Observation of multipactor suppression in a dielectric-loaded accelerating structure using an applied axial magnetic field. Appl. Phys. Lett. 103, 213503. https://doi.org/10.1063/1.4832326 Kaiser, J., Xu, C., Eichler, A., Garcia, A.S.,

    doi

  70. [2016]

    Association for Computing Machinery, New York, NY, USA, pp

    XGBoost: A Scalable Tree Boosting System, in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16. Association for Computing Machinery, New York, NY, USA, pp. 785–794. https://doi.org/10.1145/2939672.2939785 30 CST Studio Suite,

  71. [2017]

    and at the Los Alamos Neutron Science Center (LANSCE) linear accelerators, where multipactor prompted RF injector redesigns (Simakov,Evgenya et al., 2024; Xu,Haoran et al., 2024). These examples underscore the urgent need for predictive multipactor modeling to improve reliabil...

    2024

  72. [2018]

    https://doi.org/10.48550/arXiv.1811.03172 Energy, U.-B.L

    Opportunities in Machine Learning for Particle Accelerators. https://doi.org/10.48550/arXiv.1811.03172 Energy, U.-B.L. for the U.D. of,

    doi arXiv

  73. [2019]

    Special issue,

    were developed to capture the stochastic and dynamic nature of multipactor growth. PIC simulations, in particular, have become powerful tools for modeling time-dependent electron motion and space-charge effects in realistic electromagnetic fields. Nevertheless, MC, PIC, and st...

    2018

  74. [2020]

    https://doi.org/10.48550/arXiv.1907.02893 Benford, J., Schamiloglu, E., Stephens, J.C., Swegle, J.A., Zhang, P.,

    Invariant Risk Minimization. https://doi.org/10.48550/arXiv.1907.02893 Benford, J., Schamiloglu, E., Stephens, J.C., Swegle, J.A., Zhang, P.,

    1907 doi arXiv

  75. [2021]

    Fusion Eng

    An Application of Machine Learning for Plasma Current Quench Studies via Synthetic Data Generation. Fusion Eng. Des. 171, 112578. https://doi.org/10.1016/j.fusengdes.2021.112578 Edelen, A., Mayes, C., Bowring, D., Ratner, D., Adelmann, A., Ischebeck, R., Snuverink, J., Agapov,...

    2021

  76. [2022]

    CST Studio Suite,

    performed in CST Particle Studio (CST PS) (“CST Studio Suite,” 2019). The simulation domain consisted of a parallel-plate geometry, excited by a 1 GHz time-harmonic RF signal applied across a variable vacuum gap distance (d). Secondary electron yield (SEY) was modeled using Va...

    2019

  77. [2023]

    Inverse Distance Weighting to Rapidly Generate Large Simulation Datasets. J. Biomech. 158, 111764. https://doi.org/10.1016/j.jbiomech.2023.111764 32 Khamlich, M., Pichi, F., Girfoglio, M., Quaini, A., Rozza, G.,

    2023

  78. [2024]

    Results Phys

    A rapid method for prediction of the non-resonant ultra-fast multipactor regime in high gradient RF accelerating structures. Results Phys. 64, 107921. https://doi.org/10.1016/j.rinp.2024.107921 Graves, T.P. (Timothy P.,

    2024

  79. [2025]

    Optimal transport-based displacement interpolation with data augmentation for reduced order modeling of nonlinear dynamical systems. J. Comput. Phys. 531, 113938. https://doi.org/10.1016/j.jcp.2025.113938 Kishek, R.A.,

    2025

  80. [2752]

    https://doi.org/10.3390/electronics10222752 Schapire, R.E.,

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.