Pith. sign in

REVIEW 3 major objections 4 minor 30 references

Generative AI Models for Learning Flow Maps of Stochastic Dynamical Systems in Bounded Domains

T0 review · 3 major / 4 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 A hybrid machine-learning surrogate can learn stochastic flow maps in bounded domains, including where particles exit, by pairing an exit-probability network with a training-free conditional diffusion model.

desk verdict A useful hybrid surrogate for bounded-domain SDEs with a clean empirical story, but the advertised convergence guarantee does not follow from the proof as written. read the letter →

arxiv 2507.15990 v1 pith:QAXQ57OP submitted 2025-07-17 stat.ML cs.LG

Minglei Yang , Yanfang Liu , Diego del-Castillo-Negrete , Yanzhao Cao , Guannan Zhang This is my paper · ORCID
classification stat.MLcs.LG
keywords stochasticdifferentialequationsboundeddomainsexitprobabilityflowmaplearningdiffusionmodelsmachinesurrogaterunawayelectrons
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

Stochastic differential equations in bounded domains are hard to simulate because particles can hit the boundary and exit, and those exit events change the distribution everywhere, not just at the boundary. This paper proposes a single machine-learned surrogate that separates the two difficulties: a neural network trained with binary cross-entropy predicts the probability that a particle exits within the next time step, and a training-free conditional diffusion model generates the next state for particles that stay inside. The two are combined in a sampling loop that kills a trajectory with the predicted exit probability and otherwise advances it. The paper demonstrates on a 1D Brownian benchmark, a 2D cellular flow, and a 3D runaway-electron model that the resulting flow map matches Monte Carlo benchmarks and evaluates about 700 times faster at 200k samples.

What carries the argument

The load-bearing object is the one-step exit probability $P_{\mathrm{exit}}(x)$, defined as the probability that the first exit time $\theta_{n,x}$ occurs within one time step $\Delta t$ from state $x$. A fully connected sigmoid-output network $F_\eta$ approximates it via binary cross-entropy loss on exit indicators; the paper establishes uniform convergence in probability of $F_\eta$ to $P_{\mathrm{exit}}$ under Lipschitz regularity and a global-minimum assumption. The second component is a training-free conditional diffusion model: closed-form forward and reverse score functions estimated by Monte Carlo from trajectory data, converted to a deterministic reverse ODE, generate labeled state transitions used to train a supervised generator $G_\xi(x,z)$. Algorithm 1 stitches them together by drawing a uniform random number at each step and terminating the trajectory when that number is below $F_\eta(x)$.

What would settle it

Train the same architecture on 1D Brownian exit data with only 10k starting points and compare the predicted confinement fraction at T=3 against the analytical solution; if the error stops decreasing as trajectories increase or the network converges to a local minimum with visibly wrong boundary probabilities, the global-minimum-based guarantee is not what drives the observed accuracy.

Watch

Extended reading notes

Core claim

The central claim is that the stochastic flow map of an SDE in a bounded domain can be learned by decoupling exit detection from interior state propagation. The exit probability $P_{\mathrm{exit}}(x)=P\{\theta_{n,x}-t_n<\Delta t \mid X_{t_n}=x\}$, where $\theta$ is the first exit time, is learned by a fully connected network $F_\eta$ trained with binary cross-entropy on the exit indicator; the paper proves that at a global loss minimum the network output converges to $P_{\mathrm{exit}}$ uniformly in probability as the training set grows. For non-exiting particles, a training-free conditional diffusion model estimates the score of the transition distribution by Monte Carlo from trajectory data and solves a reverse ODE to produce labeled transitions, which supervise a simple generator $G_\xi(x,z)$. At generation time, the two components are combined so each step either terminates the particle with probability $F_\eta(x)$ or advances it with $G_\xi(x,z)$; numerical experiments report agreement with Monte Carlo benchmarks in all three test problems and a 700x speedup at 200k samples.

Load-bearing premise

The convergence guarantee assumes the binary cross-entropy loss reaches a global minimum of the neural network, which is not established for the network parameterization; if training only finds a local minimum, the proved convergence to the true exit probability does not apply.

Editorial extensions

If this is right

  • The trained models can be evaluated at new initial conditions and beyond the training window, because both components learn functions of state rather than memorizing trajectories.
  • Confinement fractions and exit probabilities can be obtained without solving a boundary-value PDE or running dense Monte Carlo, so parameter studies become cheap.
  • Training on trajectories without the exit/confined split produces spurious boundary accumulation; the paper shows the split is necessary for physical correctness.
  • At 200,000 samples the hybrid surrogate runs in about 0.40 seconds versus about 274 seconds for Monte Carlo, so the speedup grows with sample count because evaluation cost is nearly flat.

Reading between the lines

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

  • The exit classifier is modular: any transition generator with the right marginal behavior could replace the diffusion component, and the convergence theorem for $F_\eta$ would remain unchanged, so the construction is a template for other generative models.
  • Because evaluation cost is nearly independent of sample size, the surrogate is a natural forward model for inverse problems and uncertainty quantification, where thousands of simulations are needed.
  • The proof's global-minimum condition suggests a concrete stress test: check whether supervised training with finite data and no special initialization still delivers the predicted confinement fractions; if not, the practical guarantee may need a different argument.
  • An immediate testable extension is to time-dependent, non-autonomous SDEs, where the exit probability would depend on $t_n$ as well as $x$; the paper's framework does not cover that case explicitly.
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

3 major / 4 minor

Summary. The paper proposes a hybrid data-driven surrogate for stochastic differential equations in bounded domains, decomposing the problem into an exit-probability neural network trained with binary cross-entropy loss and a training-free conditional diffusion model that generates interior state transitions. The two components are combined in a sampling algorithm that probabilistically terminates trajectories at the boundary. The method is tested on a 1D Brownian-motion example with an analytical exit probability, a 2D advection-diffusion cellular flow, and a 3D runaway-electron model, with comparisons to Monte Carlo simulations and a claimed 700x speedup.

Significance. If the claims hold, the paper addresses a genuinely useful gap: existing stochastic flow map learners target unbounded domains and do not handle absorbing boundaries. The proposed decomposition is natural and the empirical validation is quite thorough, including an exact 1D check, distribution-level comparisons in 2D and 3D, and a public code repository with reproducibility claims. The 1D confinement rates agree with ground truth to about 1%, and the 2D and 3D marginal distributions show strong qualitative agreement. However, the paper's advertised 'rigorous convergence guarantees' are not established by the provided proof, and this weakness is load-bearing because the abstract and introduction explicitly promise this theoretical contribution.

major comments (3)
  1. [Section 3.1.2, Lemma 2 and Theorem 1] The proof of Lemma 2 is invalid. From the global-minimum condition ∇ηL_BCE(η)=0 one cannot conclude that the gradient of the local contribution ∇ηL_BCE^(r)(η,x_m) in Eq. (40) is zero; the gradient of a sum vanishing does not imply that each summand's gradient vanishes. The subsequent division by ∇ηFη(x_j) also assumes nonzero gradients without justification. Furthermore, even if Lemma 2 were repaired, the final step of Theorem 1 requires uniform control of max_{j=1,...,M} |Fη(x_j)-Pexit(x_j)|, whereas Lemma 2 only provides pointwise convergence for each fixed sample point; as M grows, the maximum over an increasing number of random points need not vanish. Therefore the claimed rigorous convergence result in Eq. (42) is unsupported.
  2. [Section 3.1.2, Assumption 1] The i.i.d. assumption on the training sample starting points is acknowledged in the text as not strictly true, since the points are produced by SDE trajectories. More seriously, the bounded-below density condition 0<ρmin≤ρ(x) cannot hold near absorbing boundaries, where surviving trajectory segments become sparse exactly in the region where Pexit is largest. This undermines Lemma 1's covering-radius argument and with it the proof of Theorem 1. The authors should either develop a convergence analysis under the actual trajectory-dependent sampling process or explicitly downgrade the result to an empirical/heuristic statement.
  3. [Section 3.1.1, Eqs. (16)-(18)] The derivation leading to Fη(x)=Jexit/J as J→∞ describes the unconstrained population minimizer of the BCE loss at a single location. In the actual training, Fη is a single network shared over all x and is subject to architecture and optimization constraints, so the stationary-point argument does not imply the trained network converges to the true Pexit(x). This is not a fatal flaw for the empirical method, but it should be stated as a heuristic motivation rather than as part of the convergence proof.
minor comments (4)
  1. [Section 3.2, Eq. (53)] The loss for training Gξ is written with a minus sign in front of the mean squared error: L_G(ξ) = -1/M Σ (y_m - Gξ(x_m,z_m))^2. Minimizing this would maximize the squared error; this is almost certainly a sign typo and should be corrected.
  2. [Section 4.4] The text refers to 'the PR-NF method' when assessing the accuracy of the runaway-electron predictions, but PR-NF is never defined in this manuscript and appears to be leftover from a previous paper.
  3. [Table 2] The evaluation times are formatted inconsistently, e.g., '39.05e-02 (secs)' and '40.46e-02 (secs)'; using a single decimal convention would improve readability.
  4. [Section 4.2] The 2D cellular flow is described as having 'chaotic advection' in the conclusion, but the flow in Eqs. (58)-(59) is steady and has regular streamlines; 'chaotic advection' is not an appropriate characterization here.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the hybrid model is a supervised learning pipeline validated against external Monte Carlo and analytic benchmarks; self-citations supply implementation details, and the convergence proof gap in Theorem 1 is a correctness issue, not a circular reduction.

full rationale

The paper's central derivation is a supervised learning consistency argument, not a circular one. The exit-probability network F_eta is trained with binary cross-entropy loss on exit indicators gamma_m drawn from the same SDE whose flow map is the target, and Eq. (18) shows that the per-location BCE stationary point is the empirical exit frequency. The claimed convergence to P_exit is a law-of-large-numbers statement, not a definitional identity. The diffusion generator G is trained by MSE on labels produced by a reverse-time ODE whose score is estimated from Sobs; this is distillation of an estimator, not prediction of the training labels as if they were independent ground truth. The 1D, 2D, and 3D validations compare against an analytic solution and against Monte Carlo simulations, i.e., external benchmarks, so the empirical claims are not forced by construction. The self-citations to [13,14] supply the training-free diffusion score-estimation details; those details are parameter-free derivations of Gaussian perturbation kernels plus Monte Carlo estimation, and the current paper states the governing forward/reverse equations, so the citation is not an unverified premise that forbids alternatives. The main weakness is the convergence proof: Lemma 2 in Section 3.1.2 assumes that the global stationarity of the full BCE loss implies stationarity of the local sub-loss in Eq. (40), which does not follow, and Assumption 1's i.i.d. sampling is acknowledged in the text as not strictly true. This is a correctness/rigor gap in Theorem 1, not circularity, because the conclusion is not equivalent to an input by construction. Accordingly, no circular step meets the evidentiary bar; score 2 reflects only the presence of self-citations and the theory's dependence on them for implementation details.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method does not introduce new physical entities or forces. It relies on standard SDE well-posedness and several empirical assumptions about the training data and neural network optimization. The free parameters are mostly standard ML hyperparameters plus the unverified i.i.d. and global optimum assumptions in the convergence proof.

free parameters (5)
  • Neural network hyperparameters for F_eta = Not fully specified in text; 3 hidden layers of 256 neurons, dropout 0.2, learning rate 0.005 (Section 4.3.2)
    The exit probability network's architecture and training hyperparameters are chosen by the authors and affect the empirical accuracy of Pexit. They are not derived from theory.
  • Neural network hyperparameters for G_xi = One hidden layer of 128 neurons, training for 5000 epochs (Section 4.4)
    The generator network's architecture and training epochs are chosen empirically and affect the fidelity of the transition distribution.
  • K, number of discretization steps for reverse ODE = 5000 (Section 4.4)
    The reverse ODE solver uses K = 5000 steps; this is a numerical discretization choice that affects accuracy.
  • Number of nearest neighbors for score estimation = 2048 (Section 4.4)
    The Monte Carlo score estimation uses 2048 nearest neighbors; this is an empirical hyperparameter.
  • Training data size and sampling interval = 50K training samples for the 3D case (Table 2); Delta T = 0.05 for 1D/2D, Delta t_obs = 0.2 for 3D
    The amount and temporal resolution of training data affect the learned models. These are choices, not derived from the theory.
assumptions (4)
  • standard math The drift a and diffusion b in Eq. (1) are globally Lipschitz, ensuring existence and uniqueness of the SDE solution.
    Standard condition for well-posedness of the SDE, invoked in Section 2 without proof.
  • domain assumption The exit probability Pexit is L-Lipschitz continuous on D, and the training sample points are i.i.d. from a density rho bounded above and below.
    Assumption 1 in Section 3.1.2. The i.i.d. assumption is acknowledged as not strictly true for trajectory data, and the Lipschitz regularity of Pexit is not verified for the test problems.
  • domain assumption The neural network F_eta has sufficient capacity to approximate any continuous function, i.e. universal approximation.
    Invoked in Section 3.1.1 to justify convergence of the trained network to Pexit. Standard but unproven for the specific architecture and training algorithm.
  • domain assumption The BCE loss converges to a global minimum, and the gradient of F_eta at sample points is nonzero.
    Lemma 2 and Theorem 1 assume the loss reaches a global minimum, which is not guaranteed for nonconvex neural network training. The nonzero gradient assumption is also unstated in the theorem statement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative AI Models for Learning Flow Maps of Stochastic Dynamical Systems in Bounded Domains." pith.science (2026). https://pith.science/paper/QAXQ57OP

@misc{pith2026250715990,
  author       = {Pith},
  title        = {Pith review of: Generative AI Models for Learning Flow Maps of Stochastic Dynamical Systems in Bounded Domains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QAXQ57OP}},
  note         = {Machine review of arXiv:2507.15990}
}
read the original abstract

Simulating stochastic differential equations (SDEs) in bounded domains, presents significant computational challenges due to particle exit phenomena, which requires accurate modeling of interior stochastic dynamics and boundary interactions. Despite the success of machine learning-based methods in learning SDEs, existing learning methods are not applicable to SDEs in bounded domains because they cannot accurately capture the particle exit dynamics. We present a unified hybrid data-driven approach that combines a conditional diffusion model with an exit prediction neural network to capture both interior stochastic dynamics and boundary exit phenomena. Our ML model consists of two major components: a neural network that learns exit probabilities using binary cross-entropy loss with rigorous convergence guarantees, and a training-free diffusion model that generates state transitions for non-exiting particles using closed-form score functions. The two components are integrated through a probabilistic sampling algorithm that determines particle exit at each time step and generates appropriate state transitions. The performance of the proposed approach is demonstrated via three test cases: a one-dimensional simplified problem for theoretical verification, a two-dimensional advection-diffusion problem in a bounded domain, and a three-dimensional problem of interest to magnetically confined fusion plasmas.

Figures

Figures reproduced from arXiv: 2507.15990 by the authors.

Figure 1
Figure 1. Convergence analysis for exit probability estimation. The plot shows the decay of KL divergence as a [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗
Figure 2
Figure 2. Comparison of trajectory simulation methods for [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Exit probability Pexit(x) with ∆T = 0.05 for the 2D stochastic transport problem. Each panel shows Pexit(x1, x2) over the domain [−π, π] × [0, 2]. Left: Our neural network prediction yielding smooth, high-fidelity results. Middle: Monte Carlo simulation with nx = 101 contour grid points using 2000 samples per grid point. Right: Monte Carlo with nx = 31 contour grid points using 1000 samples per grid point. The exit … view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Exit probability at terminal time Tmax as a function of initial position x1 with fixed x2 = 1. The plot shows exit probability calculated from Ntraj = 50, 000 trajectory samples for each initial position, i.e., exit probability = Nescape/Ntraj. Our unified method (red …
Figure 5
Figure 5. Figure 5: Distribution comparison for particles initially sampled from [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Distribution comparison for particles initially sampled from [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Escape probability cross-section comparison for runaway electron dynamics. Each row shows a projection [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]
Figure 8
Figure 8. Figure 8: Marginal PDFs of θ, p, and r at t = 10 and t = 30 for initial condition Tˆ0 = 4 (top), and Tˆ0 = 10 (bottom). The diffusion model (solid lines) shows strong agreement with Monte Carlo simulation (dashed lines) across all dimensions. To assess the accuracy of the PR-NF …
Figure 9
Figure 9. Figure 9: 2D distribution log10-pdf f(θ, p, t) at t = 10 (top), t = 20 (middle) and t = 30 (bottom) for initial condition Tˆ0 = 10. Left: Monte Carlo simulation. Right: Hybrid diffusion model. Note that t = 30 is a prediction time state beyond the training window. 5. Conclusion …
Figure 10
Figure 10. Figure 10: Comparison of runaway electron production [PITH_FULL_IMAGE:figures/full_fig_p024_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 25 canonical work pages

  1. [1]

    Archambeau, D

    C. Archambeau, D. Cornford, M. Opper, and J. Shawe-Taylor. Gaussian process approxima- tions of stochastic differential equations. InGaussian Processes in Practice, pages 1–16. PMLR, 2007

    work page 2007

  2. [2]

    Carbajal, D

    L. Carbajal, D. del Castillo-Negrete, D. Spong, S. Seal, and L. Baylor. Space dependent, full orbit effects on runaway electron dynamics in tokamak plasmas.Physics of Plasmas, 24(4), 2017

    work page 2017

  3. [3]

    X. Chen, L. Yang, J. Duan, and G. E. Karniadakis. Solving inverse stochastic problems from discrete particle observations using the fokker–planck equation and physics-informed neural networks. SIAM Journal on Scientific Computing, 43(3):B811–B830, 2021

    work page 2021

  4. [4]

    Chen and D

    Y. Chen and D. Xiu. Data-driven effective modeling of multiscale stochastic dynamical systems. arXiv preprint arXiv:2408.14821, 2024

    arXiv 2024

  5. [5]

    Chen and D

    Y. Chen and D. Xiu. Learning stochastic dynamical system via flow map operator.Journal of Computational Physics, 508:112984, 2024

    work page 2024

  6. [6]

    Chen and D

    Y. Chen and D. Xiu. Modeling unknown stochastic dynamical system subject to external excitation. arXiv preprint arXiv:2406.15747, 2024

    arXiv 2024

  7. [7]

    Friedrich, J

    R. Friedrich, J. Peinke, M. Sahimi, and M. R. R. Tabar. Approaching complexity by stochastic methods: From biological systems to turbulence.Physics Reports, 506(5):87–162, 2011

    work page 2011

  8. [8]

    D. J. Higham. An algorithmic introduction to numerical simulation of stochastic differential equations. SIAM review, 43(3):525–546, 2001

    work page 2001

Show all 30 references
  1. [9]

    D. J. Higham, X. Mao, M. Roj, Q. Song, and G. Yin. Mean exit times and the multilevel monte carlo method.SIAM/ASA Journal on Uncertainty Quantification, 1(1):2–18, 2013. 25

    2013

  2. [10]

    Kobyzev, S

    I. Kobyzev, S. J. Prince, and M. A. Brubaker. Normalizing flows: An introduction and review of current methods. IEEE transactions on pattern analysis and machine intelligence, 43(11): 3964–3979, 2020

    2020

  3. [11]

    V. I. Kolobov. Fokker–planck modeling of electron kinetics in plasmas and semiconductors. Computational materials science, 28(2):302–320, 2003

    2003

  4. [12]

    Y. Li, F. Zhao, S. Xu, J. Duan, and X. Liu. A deep learning method for computing mean exit time excited by weak gaussian noise.Nonlinear Dynamics, 112(7):5541–5554, 2024

    2024

  5. [13]

    Y. Liu, Y. Chen, D. Xiu, and G. Zhang. A training-free conditional diffusion model for learning stochastic dynamical systems.arXiv preprint arXiv:2410.03108, 2024

    2024 arXiv

  6. [14]

    Y. Liu, M. Yang, Z. Zhang, F. Bao, Y. Cao, and G. Zhang. Diffusion-model-assisted supervised learning of generative models for density estimation.Journal of Machine Learning for Modeling and Computing, 5(1), 2024

    2024

  7. [15]

    W. M. McEneaney. A curse-of-dimensionality-free numerical method for solution of certain hjb pdes. SIAM journal on Control and Optimization, 46(4):1239–1276, 2007

    2007

  8. [16]

    T. H. Nguyen, U. Simsekli, M. Gurbuzbalaban, and G. Richard. First exit time analysis of stochastic gradient descent under heavy-tailed gradient noise.Advances in neural information processing systems, 32, 2019

    2019

  9. [17]

    A. Q. Nichol and P. Dhariwal. Improved denoising diffusion probabilistic models. InInterna- tional conference on machine learning, pages 8162–8171. PMLR, 2021

    2021

  10. [18]

    M. Opper. Variational inference for stochastic differential equations.Annalen der Physik, 531 (3):1800233, 2019

    2019

  11. [19]

    A. G. Peeters and D. Strintzi. The fokker-planck equation, and its application in plasma physics. Annalen der Physik, 520(2-3):142–157, 2008

    2008

  12. [20]

    A. Pinkus. Approximation theory of the mlp model in neural networks.Acta numerica, 8: 143–195, 1999

    1999

  13. [21]

    Qi and J

    D. Qi and J. Harlim. A data-driven statistical-stochastic surrogate modeling strategy for complex nonlinear non-stationary dynamics. Journal of Computational Physics, 485:112085, 2023

    2023

  14. [22]

    K. Sobczyk. Stochastic differential equations: with applications to physics and engineering, volume 40. Springer Science & Business Media, 2013

    2013

  15. [23]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  16. [24]

    Talkner, P

    P. Talkner, P. Hänggi, E. Freidkin, and D. Trautmann. Discrete dynamics and metastability: Mean first passage times and escape rates.Journal of Statistical Physics, 48:231–254, 1987

    1987

  17. [25]

    L. Yang, D. Zhang, and G. E. Karniadakis. Physics-informed generative adversarial networks for stochastic differential equations.SIAM Journal on Scientific Computing, 42(1):A292–A317, 2020. 26

    2020

  18. [26]

    M. Yang, G. Zhang, D. del Castillo-Negrete, and M. Stoyanov. A feynman-kac based numerical method for the exit time probability of a class of transport problems.Journal of Computational Physics, 444:110564, 2021

    2021

  19. [27]

    M. Yang, D. del Castillo-Negrete, Y. Cao, and G. Zhang. A probabilistic scheme for semilinear nonlocal diffusion equations with volume constraints.SIAM Journal on Numerical Analysis, 61(6):2718–2743, 2023

    2023

  20. [28]

    M. Yang, P. Wang, D. del Castillo-Negrete, Y. Cao, and G. Zhang. A pseudoreversible normal- izing flow for stochastic dynamical systems with various initial distributions.SIAM Journal on Scientific Computing, 46(4):C508–C533, 2024

    2024

  21. [29]

    Zarzoso, D

    D. Zarzoso, D. del Castillo-Negrete, R. Lacroix, P.-E. Bernard, and S. Touzet. Transport and losses of fusion-born alpha particles in the presence of tearing modes using the new toroidal accelerated particle simulator (tapas). Plasma Physics and Controlled Fusion, 64(4):044003, 2022

    2022

  22. [30]

    Zhang and D

    G. Zhang and D. del Castillo-Negrete. A backward monte-carlo method for time-dependent runaway electron simulations.Physics of Plasmas, 24(9), 2017. 27

    2017

Pith tools

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