Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

This paper shows that a differentiable simulator can learn Fokker-Planck collision operators directly from plasma phase-space snapshots, achieving lower rollout error than particle-track statistics and matching analytical theory.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 10:10 UTC pith:ZSKAKJXN

load-bearing objection Solid inverse-problem approach with honest benchmarking; main gap is unverified Markovian/cadence assumption. the 3 major comments →

arxiv 2601.10885 v2 pith:ZSKAKJXN submitted 2026-01-15 physics.plasm-ph cs.LGphysics.comp-ph

Learning collision operators from plasma phase space data using differentiable simulators

Diogo D. Carvalho , Pablo J. Bilbao , Warren B. Mori , Luis O. Silva , E. Paulo Alves This is my paper
classification physics.plasm-ph cs.LGphysics.comp-ph
keywords collision operatorsFokker-Planckdifferentiable simulationparticle-in-cellinverse problemplasma physicsphase space diagnostics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that the advection and diffusion coefficients of a Fokker-Planck collision operator can be learned directly from phase-space snapshots of particle subpopulations, without ever storing individual particle trajectories. A differentiable Fokker-Planck solver is used to unroll the evolution of these subpopulations forward in time, and gradient-based optimization adjusts the operator parameters to minimize the mismatch with observed phase-space dynamics. On data from 76 two-dimensional particle-in-cell simulations of uniform thermal electron plasmas, the learned operators give lower rollout errors than coefficients estimated from particle tracks, and they match the theoretical finite-size PIC collision operator across shape functions, grid resolutions, particles per cell, and thermal velocities. The authors argue this opens a route to infer collision operators in regimes where analytical theory is unavailable or expected to fail.

Core claim

The central discovery is that the inverse problem of extracting a Fokker-Planck operator from kinetic data can be solved by optimizing for long-term phase-space prediction rather than for short-time velocity-change statistics. Tracking the evolution of multiple subpopulations that cover distinct velocity regions breaks the inherent non-uniqueness of fitting advection and diffusion from a single distribution, and temporal unrolling through a differentiable solver biases the recovered coefficients toward operators that reproduce the dynamics over many time steps. In the non-relativistic, electrostatic limit, the retrieved operators agree with the theoretical finite-size particle-in-cell collis

What carries the argument

The central object is the Fokker-Planck operator written as ∂_t f = -∇_v·(A f) + ½ ∇_v·[∇_v·(D f)], with advection vector A and diffusion tensor D to be learned. The machinery that carries the argument is a differentiable Fokker-Planck solver: phase-space distributions are advanced with an explicit Euler step using centered finite differences, while gradients of the rollout error with respect to A and D are backpropagated through the unrolled trajectory. To make the inverse problem well-posed, the paper uses several specially selected subpopulations (centered normals, rings, quadrants) whose joint evolution constrains the operator, longer temporal unrolling during training to improve long-te

Load-bearing premise

The load-bearing premise is that the true particle dynamics in each simulation are exactly reproduced by a single, time-independent Fokker-Planck operator acting linearly in velocity space, an assumption enforced in practice by discarding all simulations with more than 1% total-energy drift.

What would settle it

A definitive test: apply the identical pipeline to a plasma in a regime where the Fokker-Planck assumption is expected to fail — for example, a strongly coupled or relativistically hot plasma — and check whether any time-independent A and D can drive the rollout error below the level of the training data noise while remaining consistent across disjoint subpopulations. If no operator reproduces the phase-space evolution, or if the recovered coefficients depend on which subpopulations are used, the central claim that this data contains a unique FP collision operator is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Learned operators improve on particle-track estimates for long-term rollout accuracy, since they are optimized to match the phase-space dynamics at many future times rather than one-step moments.
  • The method avoids storing full particle trajectories: only phase-space snapshots are needed, substantially reducing memory costs for large 3D simulations and enabling higher-cadence diagnostics.
  • No prior knowledge of the relevant time scales is needed, because the optimization does not require choosing a statistically linear measurement interval.
  • The recovered operators match the theoretical electrostatic PIC collision operator down to small corrections, validating both the theory and the approach across shape functions, grids, particle weights, and thermal velocities.
  • Because the operator form is fixed but coefficients are free, the same pipeline can be applied to learn operators in regimes where closed-form theory is missing, such as relativistic or electromagnetically dominated collisions.

Where Pith is reading between the lines

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

  • If the approach transfers to relativistic, electromagnetically dominated plasmas, it could provide empirical collision operators for astrophysical and inertial-confinement settings where the standard Fokker-Planck coefficients are expected to break down — the authors state this as future work, but the implied payoff is direct reduced models for transport and acceleration.
  • The method doubles as a model-class probe: if a dataset cannot be reproduced by any time-independent Fokker-Planck operator with the enforced symmetries, that failure would itself diagnose non-Markovian or time-varying collision dynamics, turning the optimizer into a test of the FP hypothesis.
  • Using phase-space snapshots from experimental diagnostics (e.g., laser-scattering or radiation-belt measurements) rather than particle-in-cell data is a natural extension, since the learning pipeline never requires particle identities — though the inference would inherit the limited velocity-space coverage of the experiment.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes an inverse-problem framework for learning Fokker-Planck advection and diffusion coefficients directly from phase-space distribution data, using a differentiable Fokker-Planck solver and gradient-based optimization. The method is tested on 2D electromagnetic PIC simulations of uniform thermal plasmas, with operators parameterized as discrete tensors, neural networks, and a multi-simulation neural network. The central quantitative claim is that operators learned from phase-space subpopulation evolution achieve long-term rollout errors comparable to or better than standard particle-track estimates, while requiring only phase-space diagnostics. The paper also compares the recovered coefficients with the analytic finite-size PIC collision operator of Touati et al. (2022) for varied shape functions, grid resolutions, particle-per-cell, and thermal velocities, reporting excellent agreement. Strengths include a substantial simulation dataset, explicit rollout-error comparisons, a careful discussion of non-uniqueness, and the use of temporal unrolling and symmetry constraints.

Significance. If the central claims hold, the method is a practically useful tool for extracting interpretable collision operators from kinetic simulations without storing particle tracks, and it provides the first broad quantitative verification of the theoretical PIC collision operator in a fully electromagnetic PIC code. The paper is honest about the Markovian/time-homogeneous assumption and about the energy-filter criterion, and it ships a self-contained comparison against the existing theory. However, two load-bearing points need attention before the claims can be accepted at face value: (i) the recovered operator is tested at a single diagnostic cadence per simulation, so cadence-dependence of the learned coefficients is not assessed; and (ii) the multi-simulation neural network is evaluated on simulations used in training, so the paper does not yet demonstrate parameter-space generalization. These issues do not invalidate the per-simulation methodology, but they limit the strength of the advertised generalizations.

major comments (3)
  1. [§2.1, §3.2, §3.3, §4]
  2. [§3.2, §3.5, Supplementary S7]
  3. [Supplementary S2, Table S1, §3.5, Fig. 10]
minor comments (6)
  1. [Abstract] Typo: "wide rage" should be "wide range."
  2. [§2.2] Duplicate phrase: "advection and diffusion and diffusion coefficients" should be "advection and diffusion coefficients."
  3. [§4] Duplicate phrase: "directions for for future work" should be "directions for future work."
  4. [Figure 10] The legend in the bottom two rows uses "Stats" while the text and top row use "Tracks"; unify the nomenclature.
  5. [Eq. (3.1)] The sentence following Eq. (3.1) about the L1 norm is garbled in the text and should be typeset properly.
  6. [Supplementary S4] The notation for the number of dumped time-steps in a track-statistics interval is called both N_d and N_t in the text and in Figure S5; choose one symbol to avoid confusion with the rollout length N_t.

Circularity Check

0 steps flagged

No significant circularity: operators are fitted to external PIC phase-space data and tested on held-out subpopulations; the Touati et al. comparison is a parameter-free external check, with only a minor self-citation from overlapping authors.

full rationale

The paper's derivation chain is: (i) generate PIC phase-space data with OSIRIS (an external first-principles electromagnetic simulation), (ii) fit A,D within the Fokker-Planck ansatz (1.1) by minimizing the rollout loss (2.6) against those data, (iii) evaluate the fitted operator on held-out test subpopulations (Rings, Quadrants), and (iv) compare the retrieved operator with the Touati et al. (2022) theory (1.2-1.5). None of these steps reduces to its own inputs: the target trajectories are produced by self-consistent PIC dynamics rather than by solving the FP equation being learned; the learned A,D are not used to generate the data they are validated against; and the theory is a parameter-free analytic prediction evaluated at the externally set simulation parameters (Nppc, m, Δx/λD, vth), with its stated electrostatic/no-aliasing assumptions not including the fitted values. The A/D non-uniqueness is explicitly acknowledged (S5) and mitigated with multiple subpopulations and temporal unrolling, so the split between advection and diffusion is not fixed by construction. The claim that phase-space-learned operators beat particle-track estimates is supported by rollout MAE on training- and test-subpopulations; note that the evaluation metric coincides with the training objective and that Tracks operators are integrated at Δtdump/10, which is a fairness caveat rather than circularity. The only circularity-adjacent factor is that the validation theory (Touati et al. 2022) is co-authored by two of the present authors (Mori, Silva); however, this is not load-bearing because the theory is re-derived in S1 from stated assumptions, is parameter-free, and is externally falsifiable against the present PIC data. Disclosed limitations — discarding runs with ΔE/E > 10^-2 (S2), the high-v extrapolation caveat (S9.1), and the absence of a diagnostic-cadence invariance test — affect the interpretation of the recovered operator as a physical rather than cadence-effective object (a correctness risk), but do not constitute circular reasoning.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim rests on the FP ansatz, stationarity of the background, and test-particle independence. The only fitted parameters are the operator coefficients themselves (tensor values or NN weights) and optimization hyperparameters. No new physical entities are introduced.

free parameters (3)
  • Discrete operator tensor values (A_x, A_y, Dxx, Dyy, Dxy) = 5 × Nv×Nv values per simulation (Nv=51), e.g. ~13,000 numbers
    For PS-Tensor, the advection/diffusion coefficients are free parameters on a fixed grid, optimized to minimize the rollout loss (Eq. 2.6).
  • Neural network weights (PS-NN, PS-NN-Multi) = 2-3 hidden layers × 128 neurons, values not given
    MLP parameterizations A_i(v)=NN(v) and D_ij(v)=NN(v) (plus simulation parameters in PS-NN-Multi) are fitted by Adam; these are the model free parameters.
  • Temporal unroll length and curriculum = N_u_max = 10, curriculum stages in Table S2
    Hyperparameters chosen by the authors based on rollout-error improvement (Fig. 6); they influence the effective operator but are not physical constants.
axioms (4)
  • domain assumption Collision dynamics follow a linear Fokker-Planck equation with advection A and diffusion D (Eq. 1.1)
    The operator is assumed to be of FP form (small-angle scattering). The paper states this is valid for weakly coupled plasmas; the method does not learn more general operator forms.
  • domain assumption The operator is time-independent and spatially uniform
    Runs with energy variation >1% are discarded (S2) to ensure stationary coefficients; the plasma is spatially uniform and a single electron species is considered.
  • domain assumption Each subpopulation evolves independently under the same A,D (test-particle/linearity)
    The optimization (Eq. 2.6) sums over subpopulations s, treating each f_s as evolving under the same operator; this assumes negligible self-consistent back-reaction of the subpopulation on the background.
  • standard math The theoretical PIC collision operator (Touati et al. 2022, Eqs. 1.2-1.5) is correct for the tested electrostatic regime
    Used as ground truth for the theory comparison (Fig. 10); the derivation is published and reproduced in Supplementary S1.

pith-pipeline@v1.3.0-alltime-deepseek · 45084 in / 8605 out tokens · 101232 ms · 2026-08-03T10:10:40.606047+00:00 · methodology

0 comments
read the original abstract

We propose a methodology to infer collision operators from phase space data of plasma dynamics. Our approach combines a differentiable kinetic simulator, whose core component in this work is a differentiable Fokker-Planck solver, with a gradient-based optimisation method to learn the collisional operators that best describe the phase space dynamics. We test our method using data from two-dimensional Particle-in-Cell simulations of spatially uniform thermal plasmas, and learn the collision operator that captures the self-consistent electromagnetic interaction between finite-size charged particles over a wide variety of simulation parameters. We demonstrate that the learned operators are more accurate than alternative estimates based on particle tracks, while making no prior assumptions about the relevant time scales of the processes and significantly reducing memory requirements. We find that the retrieved operators, obtained in the non-relativistic regime, are in excellent agreement with theoretical predictions derived for electrostatic scenarios. Our results show that differentiable simulators offer a powerful and computational efficient approach to infer novel operators for a wide rage of problems, such as electromagnetically dominated collisional dynamics and stochastic wave-particle interactions.

Figures

Figures reproduced from arXiv: 2601.10885 by Diogo D. Carvalho, E. Paulo Alves, Luis O. Silva, Pablo J. Bilbao, Warren B. Mori.

Figure 1
Figure 1. Figure 1: Illustration of how advection and diffusion coefficients can be inferred from 2D particle tracks. (a) By following a group of particles with similar initial velocities over time, we observe that advection leads to an average velocity drift (< ∆vi >), while diffusion leads to an increased spread of the distribution (< ∆vi∆vj >). Drift is visible by noting that the average particle velocity (red line) is cha… view at source ↗
Figure 2
Figure 2. Figure 2: (a) Illustration of how advection and diffusion coefficients are inferred from the 2D phase space evolution of Ns subpopulations using a differentiable Fokker-Planck (FP) solver. The evolution of the phase space and the changes in the operator are exaggerated for visualization purposes. (b) Using the FP solver and the current operator state, we advance the phase space over Nu time steps and compare against… view at source ↗
Figure 3
Figure 3. Figure 3: Advection and diffusion coefficients retrieved with different approaches for simulation index = 0 (Nppc = 4, m = 1, ∆x/λD = 1, vth = 0.01c). Track operator is computed from statistics over all macroparticles, PS operators are computed from the phase space evolution of 9 sub-populations using a differentiable FP solver and a discrete (PS-Tensor) or continuous (PS-NN, PS-NN-Multi) approximator. Note that all… view at source ↗
Figure 4
Figure 4. Figure 4: Phase space evolution for a ring subpopulation using operators recovered in [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Distribution of rollout errors for advection and diffusion models obtained from particle tracks (Tracks) or phase space evolution of subpopulations (PS-Tensor, PS-NN, PS-NN-Multi). Boxplots represent statistics over the full dataset of PIC simulations (see table S1 for more details) averaged over initial subpopulations (Train - 9 subpopulations; Test - 19 subpopulations, more information in Supplementary M… view at source ↗
Figure 6
Figure 6. Figure 6: Impact of maximum temporal unroll length during training (N max u ) on the long-term rollout error of different models. Rollout length is Nt ≈ 100 across the full dataset of subpopulations and simulations. Larger temporal unrolling at train time consistently leads to improved rollout performance across all models, showcasing the importance of optimizing the operators for long-term prediction. 3.4. Enforcin… view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of the impact of maximum training temporal unroll length on PS-Tensor models. The example shown corresponds to simulation index = 0 (Nppc = 4, m = 1, ∆x/λD = 1, vth = 0.01c). It is clear that increasing N max u leads to a smoother operator defined over a larger region of the phase space while also significantly changing the average advection and diffusion values in regions of v ≈ 0. Tensor (di… view at source ↗
Figure 8
Figure 8. Figure 8: Impact of enforcing increasingly stricter symmetries on the rollout error. Introducing symmetries can lead to a slight increase in training error (removing some possible overfit), but consistently reduces the average test error and mitigates the appearance of outliers. meaningful due to lack of statistics. Additionally, the PS-NN results are not included solely for readability purposes since they are equiv… view at source ↗
Figure 9
Figure 9. Figure 9: Impact of enforcing symmetries into Tensor models. Example shown corresponds to simulation index = 0 (Nppc = 4, m = 1, ∆x/λD = 1, vth = 0.01c). Enforcing symmetries allows us to recover smoother and more accurate coefficients for a larger phase space region. Artifacts are nonetheless present at high v regions since very limited statistics are available at train time. 4. Conclusions In this paper we propose… view at source ↗
Figure 10
Figure 10. Figure 10: Comparison between theoretical values for advection / diffusion coefficients and those obtained from particle tracks and the PS-Tensor method for different shape functions and grid resolutions. All simulations shown use Nppc = 25 and vth = 0.01c. There is overall an excellent agreement with theory, particularly up to v/vth = 3. PS-Tensor is expected to not correctly capture well values above this threshol… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Learning time-dependent and integro-differential collision operators from plasma phase space data using differentiable simulators

    physics.plasm-ph 2026-01 unverdicted novelty 6.0

    Differentiable simulators recover time-dependent integro-differential collision operators from PIC-generated plasma data that reproduce phase-space evolution more accurately than particle-track statistics.

Reference graph

Works this paper leans on

87 extracted references · 9 linked inside Pith · cited by 1 Pith paper

  1. [1]

    In Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2\/ , pp

    Ansel, Jason , Yang, Edward , He, Horace , Gimelshein, Natalia , Jain, Animesh , Voznesensky, Michael , Bao, Bin , Bell, Peter , Berard, David , Burovski, Evgeni & others 2024 Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In Proceedings of the 29th ACM International Conference on Architectural Sup...

    2024

  2. [2]

    Oxford University Press

    Atzeni, Stefano & Meyer-ter Vehn, J \"u rgen 2004 The Physics of Inertial Fusion: BeamPlasma Interaction, Hydrodynamics, Hot Dense Matter\/ . Oxford University Press

    2004

  3. [3]

    Journal of Computational Physics 363 , 178--199

    Averkin, Sergey N & Gatsonis, Nikolaos A 2018 A parallel electrostatic particle-in-cell method on unstructured tetrahedral grids for large-scale bounded collisionless plasma simulations . Journal of Computational Physics 363 , 178--199

    2018

  4. [4]

    Physics of Plasmas 19 (3)

    Baalrud, Scott D 2012 Transport coefficients in strongly coupled plasmas . Physics of Plasmas 19 (3)

    2012

  5. [5]

    Physics of Plasmas 21 (5)

    Baalrud, Scott D & Daligault, J \'e r \^o me 2014 Extending plasma transport theory to strong coupling through the concept of an effective interaction potential . Physics of Plasmas 21 (5)

    2014

  6. [6]

    Plasma Physics and Controlled Fusion 48 (3), R37

    Bell, AR , Robinson, APL , Sherlock, M , Kingham, RJ & Rozmus, W 2006 Fast electron transport in laser-produced plasmas and the kalos code for solution of the vlasov--fokker--planck equation . Plasma Physics and Controlled Fusion 48 (3), R37

    2006

  7. [7]

    Plasma Physics and Controlled Fusion 66 (3), 035014

    Bell, AR & Sherlock, M 2024 The fastvfp code for solution of the vlasov--fokker--planck equation . Plasma Physics and Controlled Fusion 66 (3), 035014

    2024

  8. [8]

    Physics of Plasmas 32 (3)

    Bergeson, Scott , Schlitters, Matthew , Miller, Matthew , Farley, Ben , Sieverts, Devin , Murillo, Michael S & Haack, Jeffrey R 2025 Experimental and computational study of phase space dynamics in strongly coupled plasmas with steep density gradients . Physics of Plasmas 32 (3)

    2025

  9. [9]

    CRC press

    Birdsall, Charles K & Langdon, A Bruce 2018 Plasma physics via computer simulation\/ . CRC press

    2018

  10. [10]

    Naval Research Laboratory

    Boris, Jay P & Shanny, Ramy A 1972 Proceedings [of The] 4th Conference on Numerical Simulation of Plasmas\/ . Naval Research Laboratory

    1972

  11. [11]

    Physical Review Letters 59 (16), 1817

    Braams, Bastiaan J & Karney, Charles FF 1987 Differential form of the collision integral for a relativistic plasma . Physical Review Letters 59 (16), 1817

    1987

  12. [12]

    arXiv preprint arXiv:2202.03376

    Brandstetter, Johannes , Worrall, Daniel & Welling, Max 2022 Message passing neural pde solvers . arXiv preprint arXiv:2202.03376

    Pith/arXiv arXiv 2022

  13. [13]

    Proceedings of the National Academy of Sciences 113 (15), 3932--3937

    Brunton, Steven L , Proctor, Joshua L & Kutz, J Nathan 2016 Discovering governing equations from data by sparse identification of nonlinear dynamical systems . Proceedings of the National Academy of Sciences 113 (15), 3932--3937

    2016

  14. [14]

    Journal of Geophysical Research: Space Physics 127 (7), e2022JA030377

    Camporeale, Enrico , Wilkie, George J , Drozdov, Alexander Y & Bortnik, Jacob 2022 Data-driven discovery of fokker-planck equation for the earth's radiation belts electrons using physics-informed neural networks . Journal of Geophysical Research: Space Physics 127 (7), e2022JA030377

    2022

  15. [15]

    Machine Learning: Science and Technology 5 (2), 025048

    Carvalho, Diogo D , Ferreira, Diogo R & Silva, Luis O 2024 Learning the dynamics of a one-dimensional plasma model with graph neural networks . Machine Learning: Science and Technology 5 (2), 025048

    2024

  16. [16]

    Journal of Physics A: Mathematical and General 39 (17), 4411

    Chabrier, Gilles , Saumon, D & Potekhin, AY 2006 Dense plasmas in astrophysics: from giant planets to neutron stars . Journal of Physics A: Mathematical and General 39 (17), 4411

    2006

  17. [17]

    SIAM Journal on Scientific Computing 43 (3), B811--B830

    Chen, Xiaoli , Yang, Liu , Duan, Jinqiao & Karniadakis, George Em 2021 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

  18. [18]

    Journal of Computational Physics 492 , 112400

    Chung, Kyoungseoun , Fei, Fei , Gorji, M Hossein & Jenny, Patrick 2023 Data-driven stochastic particle scheme for collisional plasma simulations . Journal of Computational Physics 492 , 112400

    2023

  19. [19]

    Computer Physics Communications 135 (2), 144--153

    Esirkepov, T Zh 2001 Exact charge conservation scheme for particle-in-cell simulation with an arbitrary form-factor . Computer Physics Communications 135 (2), 144--153

    2001

  20. [20]

    In Computational Science—ICCS 2002: International Conference Amsterdam, The Netherlands, April 21--24, 2002 Proceedings, Part III 2\/ , pp

    Fonseca, Ricardo A , Silva, Luis O , Tsung, Frank S , Decyk, Viktor K , Lu, Wei , Ren, Chuang , Mori, Warren B , Deng, Shaogui , Lee, Shiyoun , Katsouleas, T & others 2002 Osiris: A three-dimensional, fully relativistic particle in cell code for modeling plasma based accelerators. In Computational Science—ICCS 2002: International Conference Amsterdam, The...

    2002

  21. [21]

    Academic press

    Fried, Burton D & Conte, Samuel D 1961 The plasma dispersion function: the Hilbert transform of the Gaussian\/ . Academic press

    1961

  22. [22]

    Journal of Computational Physics 228 (10), 3742--3761

    Gatsonis, Nikolaos A & Spirkin, Anton 2009 A three-dimensional electrostatic particle-in-cell methodology on unstructured delaunay--voronoi grids . Journal of Computational Physics 228 (10), 3742--3761

    2009

  23. [23]

    Journal of Fluid Mechanics 996 , A25

    Gurevich, Daniel R , Golden, Matthew R , Reinbold, Patrick AK & Grigoriev, Roman O 2024 Learning fluid physics from highly turbulent data using sparse physics-informed discovery of empirical relations (spider) . Journal of Fluid Mechanics 996 , A25

    2024

  24. [24]

    Journal of Computational Physics 8 (1), 19--44

    Hockney, RW 1971 Measurements of collision and heating times in a two-dimensional thermal computer plasma . Journal of Computational Physics 8 (1), 19--44

    1971

  25. [25]

    Mathematics 9 (12), 1384

    Holloway, Ian , Wood, Aihua & Alekseenko, Alexander 2021 Acceleration of boltzmann collision integral calculation using machine learning . Mathematics 9 (12), 1384

    2021

  26. [26]

    Journal of Computational Physics 419 , 109665

    Hwang, Hyung Ju , Jang, Jin Woo , Jo, Hyeontae & Lee, Jae Yong 2020 Trend to equilibrium for the kinetic fokker-planck equation via the neural network approach . Journal of Computational Physics 419 , 109665

    2020

  27. [27]

    arXiv preprint arXiv:2410.12683

    Joglekar, Archis S 2024 Generative neural reparameterization for differentiable pde-constrained optimization . arXiv preprint arXiv:2410.12683

    Pith/arXiv arXiv 2024

  28. [28]

    Journal of Plasma Physics 88 (6), 905880608

    Joglekar, Archis S & Thomas, Alexander GR 2022 Unsupervised discovery of nonlinear plasma physics using differentiable kinetic simulations . Journal of Plasma Physics 88 (6), 905880608

    2022

  29. [29]

    Machine Learning: Science and Technology 4 (3), 035049

    Joglekar, Archis S & Thomas, Alexander GR 2023 Machine learning of hidden variables in multiscale fluid simulation . Machine Learning: Science and Technology 4 (3), 035049

    2023

  30. [30]

    Physics of Plasmas 31 (2)

    Jubin, Sierra , Powis, Andrew Tasman , Villafana, Willca , Sydorenko, Dmytro , Rauf, Shahid , Khrabrov, Alexander V , Sarwar, Salman & Kaganovich, Igor D 2024 Numerical thermalization in 2d pic simulations: Practical estimates for low-temperature plasma simulations . Physics of Plasmas 31 (2)

    2024

  31. [31]

    arXiv preprint arXiv:2205.04611

    Karnakov, Petr , Litvinov, Sergey & Koumoutsakos, Petros 2022 Optimizing a discrete loss (odil) to solve forward and inverse problems for partial differential equations using machine learning tools . arXiv preprint arXiv:2205.04611

    Pith/arXiv arXiv 2022

  32. [32]

    Japanese Journal of Applied Physics 61 (8), 086002

    Kawaguchi, Satoru & Murakami, Tomoyuki 2022 Physics-informed neural networks for solving the boltzmann equation of the electron velocity distribution function in weakly ionized plasmas . Japanese Journal of Applied Physics 61 (8), 086002

    2022

  33. [33]

    Physics Reports 449 (4-5), 77--130

    Killian, Thomas C , Pattard, T , Pohl, T & Rost, JM 2007 Ultracold neutral plasmas . Physics Reports 449 (4-5), 77--130

    2007

  34. [34]

    arXiv preprint arXiv:1412.6980

    Kingma, Diederik P 2014 Adam: A method for stochastic optimization . arXiv preprint arXiv:1412.6980

    Pith/arXiv arXiv 2014

  35. [35]

    Journal of Machine Learning Research 24 (89), 1--97

    Kovachki, Nikola , Li, Zongyi , Liu, Burigede , Azizzadenesheli, Kamyar , Bhattacharya, Kaushik , Stuart, Andrew & Anandkumar, Anima 2023 Neural operator: Learning maps between function spaces with applications to pdes . Journal of Machine Learning Research 24 (89), 1--97

    2023

  36. [36]

    In 73rd Annual Gaseous Electronics Virtual Conference\/

    Kushner, Mark J 2020 Plasma science 2020 decadal survey: Overview and opportunities for the low temperature plasma community. In 73rd Annual Gaseous Electronics Virtual Conference\/ . APS

    2020

  37. [37]

    Landau, Lev D 1936 Die kinetische gleichung fuer den fall coulombscher wechselwirkung . Phys. Z. Sowjet. 10 (1936) 10 , 163

    1936

  38. [38]

    In Proceedings of the fourth conference on numerical simulation of plasmas\/ , pp

    Langdon, A Bruce 1970 Nonphysical modifications to oscillations, fluctuations, and collisions due to space-time differencing. In Proceedings of the fourth conference on numerical simulation of plasmas\/ , pp. 467--495

    1970

  39. [39]

    The Physics of Fluids 13 (8), 2115--2122

    Langdon, A Bruce & Birdsall, Charles K 1970 Theory of plasma simulation using finite-size particles . The Physics of Fluids 13 (8), 2115--2122

    1970

  40. [40]

    Journal of Computational Physics 480 , 112031

    Lee, Jae Yong , Jang, Juhi & Hwang, Hyung Ju 2023 oppinn: Physics-informed neural network with operator learning to approximate solutions to the fokker-planck-landau equation . Journal of Computational Physics 480 , 112031

    2023

  41. [41]

    ESAIM: Mathematical Modelling and Numerical Analysis 55 (5), 1803--1846

    Lee, Jae Yong , Jang, Jin Woo & Hwang, Hyung Ju 2021 The model reduction of the vlasov--poisson--fokker--planck system to the poisson--nernst--planck system via the deep neural network approach . ESAIM: Mathematical Modelling and Numerical Analysis 55 (5), 1803--1846

    2021

  42. [42]

    arXiv preprint arXiv:2504.20408

    Lee, Jae Yong , Jung, Gwang Jae , Lim, Byung Chan & Hwang, Hyung Ju 2025 Fourierspecnet: Neural collision operator approximation inspired by the fourier spectral method for solving the boltzmann equation . arXiv preprint arXiv:2504.20408

    arXiv 2025

  43. [43]

    arXiv preprint arXiv:2402.16613

    Lee, Jae Yong , Schotth \"o fer, Steffen , Xiao, Tianbai , Krumscheid, Sebastian & Frank, Martin 2024 Structure-preserving operator learning: Modeling the collision operator of kinetic equations . arXiv preprint arXiv:2402.16613

    Pith/arXiv arXiv 2024

  44. [44]

    arXiv preprint arXiv:2010.08895

    Li, Zongyi , Kovachki, Nikola , Azizzadenesheli, Kamyar , Liu, Burigede , Bhattacharya, Kaushik , Stuart, Andrew & Anandkumar, Anima 2020 Fourier neural operator for parametric partial differential equations . arXiv preprint arXiv:2010.08895

    Pith/arXiv arXiv 2020

  45. [45]

    SIAM Journal on Scientific Computing 46 (2), C186--C215

    Li, Zhengyi , Wang, Yanli , Liu, Hongsheng , Wang, Zidong & Dong, Bin 2024 Solving the boltzmann equation with a neural sparse representation . SIAM Journal on Scientific Computing 46 (2), C186--C215

    2024

  46. [46]

    Computer Methods in Applied Mechanics and Engineering 433 , 117441

    List, Bjoern , Chen, Li-Wei , Bali, Kartik & Thuerey, Nils 2025 Differentiability in unrolled training of neural physics simulators on transient dynamics . Computer Methods in Applied Mechanics and Engineering 433 , 117441

    2025

  47. [47]

    Journal of Computational Physics 447 , 110676

    Lou, Qin , Meng, Xuhui & Karniadakis, George Em 2021 Physics-informed neural networks for solving forward and inverse flow problems via the boltzmann-bgk formulation . Journal of Computational Physics 447 , 110676

    2021

  48. [48]

    arXiv preprint arXiv:1910.03193

    Lu, Lu , Jin, Pengzhan & Karniadakis, George Em 2019 Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators . arXiv preprint arXiv:1910.03193

    Pith/arXiv arXiv 2019

  49. [49]

    The Physics of Fluids 18 (12), 1740--1747

    Matsuda, Y & Okuda, H 1975 Collisions in multi-dimensional plasma simulations . The Physics of Fluids 18 (12), 1740--1747

    1975

  50. [50]

    Physics of Plasmas 30 (9)

    McDevitt, Christopher J 2023 A physics-informed deep learning model of the hot tail runaway electron seed . Physics of Plasmas 30 (9)

    2023

  51. [51]

    Physics of Plasmas 32 (4)

    McDevitt, Christopher J , Arnaud, Jonathan S & Tang, Xian-Zhu 2025 a\/ An efficient surrogate model of secondary electron formation and evolution . Physics of Plasmas 32 (4)

    2025

  52. [52]

    Physics of Plasmas 32 (4)

    McDevitt, Christopher J , Arnaud, Jonathan S & Tang, Xian-Zhu 2025 b\/ A physics-constrained deep learning treatment of runaway electron dynamics . Physics of Plasmas 32 (4)

    2025

  53. [53]

    Physics Today 57 (7), 32--38

    Merlino, Robert L & Goree, John A 2004 Dusty plasmas in the laboratory, industry, and space . Physics Today 57 (7), 32--38

    2004

  54. [54]

    Journal of Plasma Physics 87 (2), 905870211

    Miller, Marco Andres , Churchill, Randy Michael , Dener, Alp , Chang, Choong-Seock , Munson, Todd & Hager, Robert 2021 Encoder--decoder neural network for solving the nonlinear fokker--planck--landau collision operator in xgc . Journal of Plasma Physics 87 (2), 905870211

    2021

  55. [55]

    Journal of Computational Physics 470 , 111541

    Miller, Sean T , Roberts, Nathan V , Bond, Stephen D & Cyr, Eric C 2022 Neural-network based collision operators for the boltzmann equation . Journal of Computational Physics 470 , 111541

    2022

  56. [56]

    Journal of Computational Physics 523 , 113665

    Noh, Hyeongjun , Lee, Jimin & Yoon, Eisung 2025 Fpl-net: A deep learning framework for solving the nonlinear fokker--planck--landau collision operator for anisotropic temperature relaxation . Journal of Computational Physics 523 , 113665

    2025

  57. [57]

    SIAM Journal on Scientific Computing 47 (2), C451--C474

    Oh, Jaemin , Cho, Seung Yeon , Yun, Seok-Bae , Park, Eunbyung & Hong, Youngjoon 2025 Separable physics-informed neural networks for solving the bgk model of the boltzmann equation . SIAM Journal on Scientific Computing 47 (2), C451--C474

    2025

  58. [58]

    The Physics of Fluids 13 (8), 2123--2134

    Okuda, Hideo & Birdsall, Charles K 1970 Collisions in a plasma of finite-size particles . The Physics of Fluids 13 (8), 2123--2134

    1970

  59. [59]

    In DPP 2025\/

    Park, Ryan Minard , Moore, Christopher H & Baalrud, Scott D 2025 Numerical thermalization in 3d pic simulations. In DPP 2025\/ . APS

    2025

  60. [60]

    Physical Review E 93 (5), 053208

    Pike, OJ & Rose, SJ 2016 Transport coefficients of a relativistic plasma . Physical Review E 93 (5), 053208

    2016

  61. [61]

    , Everson, Erik T

    Plasmapy, Community , Murphy, Nicholas A. , Everson, Erik T. , Stańczak-Marikin, Dominik , Heuer, Peter V. , Kozlowski, Pawel M. , Johnson, Elliot T. , Malhotra, Ritiek , Schaffner, David A. , Vincena, Stephen T. , Abler, Mel , Addison, James , Ahamed, Ataf Fazledin , Alarcon, Paula Valentina , Antognetti, Benjamin , Arran, Christopher , Bagherianlemraski...

    2024

  62. [62]

    arXiv preprint arXiv:2001.04385

    Rackauckas, Christopher , Ma, Yingbo , Martensen, Julius , Warner, Collin , Zubov, Kirill , Supekar, Rohit , Skinner, Dominic , Ramadhan, Ali & Edelman, Alan 2020 Universal differential equations for scientific machine learning . arXiv preprint arXiv:2001.04385

    Pith/arXiv arXiv 2020

  63. [63]

    Journal of Computational Physics 378 , 686--707

    Raissi, Maziar , Perdikaris, Paris & Karniadakis, George E 2019 Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations . Journal of Computational Physics 378 , 686--707

    2019

  64. [64]

    Physics of Plasmas 4 (5), 1286--1296

    Reynolds, Mark Anthony , Fried, BD & Morales, GJ 1997 Velocity--space drag and diffusion in a model, two-dimensional plasma . Physics of Plasmas 4 (5), 1286--1296

    1997

  65. [65]

    Physical Review 107 (1), 1

    Rosenbluth, Marshall N , MacDonald, William M & Judd, David L 1957 Fokker-planck equation for an inverse-square force . Physical Review 107 (1), 1

    1957

  66. [66]

    Science Advances 3 (4), e1602614

    Rudy, Samuel H , Brunton, Steven L , Proctor, Joshua L & Kutz, J Nathan 2017 Data-driven discovery of partial differential equations . Science Advances 3 (4), e1602614

    2017

  67. [67]

    Journal of Computational Physics 231 (3), 1051--1079

    Thomas, Alexander GR , Tzoufras, M , Robinson, APL , Kingham, Robert J , Ridgers, CP , Sherlock, Mark & Bell, AR 2012 A review of vlasov--fokker--planck numerical modeling of inertial confinement fusion plasma . Journal of Computational Physics 231 (3), 1051--1079

    2012

  68. [68]

    Plasma Physics and Controlled Fusion 64 (11), 115014

    Touati, Micha \"e l , Codur, Romain , Tsung, Frank , Decyk, Viktor K , Mori, Warren B & Silva, Luis O 2022 Kinetic theory of particle-in-cell simulation plasma and the ensemble averaging technique . Plasma Physics and Controlled Fusion 64 (11), 115014

    2022

  69. [69]

    Journal of Computational Physics 230 (17), 6475--6494

    Tzoufras, M , Bell, AR , Norreys, PA & Tsung, Frank S 2011 A vlasov--fokker--planck code for high energy density physics . Journal of Computational Physics 230 (17), 6475--6494

    2011

  70. [70]

    Physics of Plasmas 20 (5)

    Tzoufras, M , Tableman, A , Tsung, FS , Mori, WB & Bell, AR 2013 A multi-dimensional vlasov-fokker-planck code for arbitrarily anisotropic high-energy-density plasmas . Physics of Plasmas 20 (5)

    2013

  71. [71]

    Neural Networks p

    Wang, Taorui , Hu, Zheyuan , Kawaguchi, Kenji , Zhang, Zhongqiang & Karniadakis, George Em 2025 Tensor neural networks for high-dimensional fokker-planck equations . Neural Networks p. 107165

    2025

  72. [72]

    The Astrophysical Journal Letters 893 (1), L7

    Wong, Kai , Zhdankin, Vladimir , Uzdensky, Dmitri A , Werner, Gregory R & Begelman, Mitchell C 2020 First-principles demonstration of diffusive-advective particle acceleration in kinetic simulations of relativistic plasma turbulence . The Astrophysical Journal Letters 893 (1), L7

    2020

  73. [73]

    arXiv preprint arXiv:2502.03042

    Wong, Kai W , Zhdankin, Vladimir , Uzdensky, Dmitri A , Werner, Gregory R & Begelman, Mitchell C 2025 Energy diffusion and advection coefficients in kinetic simulations of relativistic plasma turbulence . arXiv preprint arXiv:2502.03042

    arXiv 2025

  74. [74]

    Journal of Computational Physics 443 , 110521

    Xiao, Tianbai & Frank, Martin 2021 Using neural networks to accelerate the solution of the boltzmann equation . Journal of Computational Physics 443 , 110521

    2021

  75. [75]

    Journal of Computational Physics 490 , 112317

    Xiao, Tianbai & Frank, Martin 2023 Relaxnet: A structure-preserving neural network to approximate the boltzmann collision operator . Journal of Computational Physics 490 , 112317

    2023

  76. [76]

    Chaos: An Interdisciplinary Journal of Nonlinear Science 30 (1)

    Xu, Yong , Zhang, Hao , Li, Yongge , Zhou, Kuang , Liu, Qi & Kurths, J \"u rgen 2020 Solving fokker-planck equation using deep learning . Chaos: An Interdisciplinary Journal of Nonlinear Science 30 (1)

    2020

  77. [77]

    IEEE Transactions on antennas and propagation 14 (3), 302--307

    Yee, Kane 1966 Numerical solution of initial boundary value problems involving maxwell's equations in isotropic media . IEEE Transactions on antennas and propagation 14 (3), 302--307

    1966

  78. [78]

    Zaharia, Matei A. , Chen, Andrew , Davidson, Aaron , Ghodsi, Ali , Hong, Sue Ann , Konwinski, Andy , Murching, Siddharth , Nykodym, Tomas , Ogilvie, Paul , Parkhe, Mani , Xie, Fen & Zumar, Corey 2018 Accelerating the Machine Learning Lifecycle with MLflow . IEEE Data Eng. Bull. 41 , 39--45

    2018

  79. [79]

    In Mathematical and scientific machine learning\/ , pp

    Zhai, Jiayu , Dobson, Matthew & Li, Yao 2022 A deep learning method for solving fokker-planck equations. In Mathematical and scientific machine learning\/ , pp. 568--597 . PMLR

    2022

  80. [80]

    International Journal of Non-Linear Mechanics 147 , 104202

    Zhang, Yang & Yuen, Ka-Veng 2022 Physically guided deep learning solver for time-dependent fokker--planck equation . International Journal of Non-Linear Mechanics 147 , 104202

    2022

Showing first 80 references.