Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

The Artificial Scientist -- in-transit Machine Learning of Plasma Simulations

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper demonstrates that plasma simulation data can be streamed directly into a machine-learning model in-transit, bypassing the filesystem, and used to learn correlations from a Kelvin-Helmholtz instability on the fly.

desk verdict A genuinely useful systems integration paper whose headline scaling claim outruns its evidence: full-Frontier streaming was shown with a no-op consumer, while the real ML pipeline stops at 96 nodes with admitted backend limits. read the letter →

arxiv 2501.03383 v3 pith:DLNR4TN7 submitted 2025-01-06 physics.comp-ph cs.DCcs.LG

Jeffrey Kelling , Vicente Bolea , Michael Bussmann , Ankush Checkervarty , Alexander Debus , Jan Ebert , Greg Eisenhauer , Vineeth Gutta
show 14 more authors
Stefan Kesselheim Scott Klasky Vedhas Pandit Richard Pausch Norbert Podhorszki Franz Poschel David Rogers Jeyhun Rustamov Steve Schmerler Ulrich Schramm Klaus Steiniger Rene Widera Anna Willmann Sunita Chandrasekaran
This is my paper · ORCID
classification physics.comp-phcs.DCcs.LG PACS 52.65.Rr
keywords in-transitmachinelearningplasmasimulationKelvin-Helmholtzinstabilityparticle-in-cellstreamingI/OcontinualopenPMDADIOS2
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 demonstrates a way to train a machine-learning model directly on the output of an exascale plasma simulation as the data is produced, never writing it to disk. The authors couple a GPU-accelerated particle-in-cell code to a machine-learning application through a streaming data interface, and use the setup to learn an inverse map from emitted radiation back to particle dynamics in a Kelvin-Helmholtz instability. They argue that this in-transit training pattern solves the filesystem bottleneck that would otherwise make petabyte-scale simulation data impossible to analyse, and that it supports continual learning from a non-steady process. As a proof of concept, the trained model reproduces radiation spectra from particle data and partially reconstructs momentum distributions, including identifying vortex regions without supervision.

What carries the argument

The central machinery is a loosely coupled in-transit data pipeline built on the openPMD data standard and the ADIOS2 streaming library, which moves particle and radiation data from the simulation's memory directly into the machine-learning application's memory over the network. The machine-learning model itself combines a transposition-invariant point-cloud encoder, a decoder, and four Glow-style coupling blocks forming an invertible neural network, trained with a five-term loss that includes Chamfer distance, KL divergence, mean-squared error, and maximum mean discrepancy terms. A separate training buffer implements experience replay, keeping a small set of recent samples alongside a random replay buffer so the model can train continually without catastrophic forgetting as the simulation evolves.

What would settle it

Run the complete PIConGPU plus machine-learning training pipeline on 9126 Frontier nodes and measure the per-node training throughput; if it falls below the 1.9-3.3 GB/s per node achieved by the no-op streaming benchmark, or if the N/RCCL communication backend fails beyond 100 nodes, the claim that the workflow scales to the full Top-1 system is contradicted.

Watch

Extended reading notes

Core claim

The paper claims that a physics simulation can be coupled to a deep-learning model so that simulation data streams directly into the training loop, with no intermediate write to storage, and that this enables learning correlations from the simulation on-the-fly. Using the relativistic Kelvin-Helmholtz instability simulated with the particle-in-cell code PIConGPU, the authors stream particle positions, momenta, and radiation spectra into an autoencoder-plus-invertible-neural-network architecture. The model learns, in an unsupervised manner, a latent representation that separates physically distinct plasma regions and a conditional inverse mapping from radiation spectra to particle momentum distributions. The authors report that, for the bulk plasma, the predicted momentum distribution agrees well with the simulation, and that the network correctly reproduces the Doppler-shifted radiation spectrum and identifies vortex regions even though the inversion is ill-posed.

Load-bearing premise

The load-bearing premise is that the full-system streaming performance measured with a synthetic no-op consumer, which does no computation, transfers to the real machine-learning training loop that must also compute gradients and synchronize across ranks.

Editorial extensions

If this is right

  • Simulations that produce more data than a filesystem can store or write can still be used for deep learning, because the data flows through memory and network rather than disk.
  • Continual-learning methods with experience replay can keep a model trained on a non-stationary simulation stream, retaining knowledge of earlier time steps while adapting to later ones.
  • The same openPMD/ADIOS2 loose-coupling pattern can be adapted to other high-rate data sources, such as high-repetition-rate detectors, where storing raw events is impossible.
  • At full system scale on Frontier, the streaming layer reaches an aggregate throughput of 20-30 TB/s, exceeding the parallel filesystem's roughly 10 TB/s bandwidth and making full-system data streaming feasible for the I/O path.
  • The model's unsupervised latent space separates approaching, receding, and vortex plasma regions, enabling classification of physical regimes without labeled training data.

Reading between the lines

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

  • The real training pipeline has only been demonstrated up to 96 nodes, while the no-op streaming benchmark reaches 9126 nodes, so the claim that the full workflow scales to Top-1 systems depends on the unverified assumption that real training can sustain the no-op throughput.
  • If the reported N/RCCL socket limit beyond 100 nodes can be circumvented, for example by using an MPI or libfabric communication backend, the same architecture could plausibly extend in-transit training to thousands of nodes.
  • The workflow suggests a general recipe for experiments with irreversible data loss: train directly on the live data stream from a detector or simulation instead of attempting to store it for offline analysis.
  • A HIP port of the KeOps library would allow the use of earth mover's distance as a loss on AMD GPUs, potentially improving the fidelity of point-cloud reconstructions beyond what Chamfer distance achieves.
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 / 5 minor

Summary. The paper presents a streaming workflow, dubbed the Artificial Scientist, that couples a PIConGPU particle-in-cell simulation of the Kelvin-Helmholtz instability to a PyTorch-based machine learning application through the openPMD/ADIOS2 in-transit I/O stack, avoiding intermediate filesystem writes. The ML model is a variational autoencoder plus invertible neural network trained with an experience-replay buffer while data streams in. The authors report PIConGPU scaling to the full 9,126-node Frontier system, a streaming benchmark with a no-op consumer achieving 20-30 TB/s at full scale, and a weak-scaling study of the coupled training pipeline from 8 to 96 nodes with roughly 35% efficiency at the largest size. The scientific evaluation is a qualitative comparison of radiation-to-momentum inversion on one selected sub-volume.

Significance. If the central claims are supported, the paper would be a valuable proof-of-concept for in-transit training of deep models on data streams too large to store, and the engineering lessons on Frontier are of broad interest to the HPC and scientific-ML communities. The paper is commendably candid about several limitations, including the hard N/RCCL socket limit, the loss of scaling efficiency, and the difficulty of transferring hyperparameters to large batch sizes. The streaming throughput measurements with the no-op consumer provide a useful scaling reference for the ADIOS2/openPMD layer, and the authors explicitly distinguish the synthetic benchmark from the real pipeline. However, the title-level claim of end-to-end scalability to Top-1 supercomputers and the qualitative-only ML evaluation mean the paper's strongest statements exceed what the data demonstrate.

major comments (3)
  1. [Section IV-B, V-A, IV-D, Abstract, Conclusion] The full-system streaming claim cited in the Abstract ('completely circumventing the capacity-constrained filesystem bottleneck') and the Conclusion ('scalable from local clusters to Top-1 supercomputers') is supported only for a no-op consumer benchmark, not for the coupled PIConGPU+MLapp pipeline. Section IV-B states the full-scale runs stream into a 'synthetic no-op consumer that performs no computation besides measuring the performance of this I/O operation and only discards received data.' The actual training pipeline is measured only up to 96 nodes (Section V-A, Fig. 8), where efficiency is about 35%, and Section IV-D explicitly states that the N/RCCL backend 'hits system limitations on the possible number of open sockets beyond 100 nodes.' Therefore the 20-30 TB/s figure and the 'Top-1 supercomputers' phrasing overstate what has been demonstrated for the workflow as a whole. Please revise the abstract and conclusion to distinguish the I/O-layer benchmark from the end-to-end training workflow, and state explicitly that the coupled pipeline was demonstrated only to 96 nodes in this study.
  2. [Section V-B, Fig. 9] The evaluation of the ML inversion, which underpins the conclusion that the model 'learn[s] correlations from a physics simulation on-the-fly,' is qualitative and based on a single 'selected example sub-volume' (Fig. 9). No quantitative metric is reported: there is no reconstruction error on held-out volumes, no classification accuracy for the claimed region identification, no comparison against a baseline or oracle, and no mention of a train/test split across time steps or spatial regions. Because the model is trained and evaluated on data from the same simulation stream, the reader cannot judge whether the model generalizes or merely memorizes. Please add quantitative, held-out evaluation (for example, Chamfer distance or density error on held-out sub-volumes and time steps) and, if the region-classification claim is retained, report a classifier accuracy with confidence intervals.
  3. [Section IV-C, experience replay] The continual-learning component is a key claimed contribution, but the paper provides no experimental evidence that the proposed replay buffer prevents catastrophic forgetting or improves on a baseline without replay. The buffer sizes and sampling counts (Nnow=10, NEP=20, nnow=4, nEP=4) are introduced as fixed choices, and the loss weights in Eq. (1) are described as empirically tuned. Given that Section V-A admits hyperparameters do not transfer from small to large scale, a sensitivity study or at least an ablation of the replay mechanism would be needed to support the claim that this scheme is effective for in-transit continual learning. Please include such an analysis or temper the claim to reflect that the replay design is a heuristic whose benefit is not demonstrated.
minor comments (5)
  1. [Section V-A] The phrase 'training with a batch size of nnow + nrep = 8 per GCD' appears to be a typo: the batch size was defined in Section IV-C as nnow + nEP = 8, while nrep is the number of training iterations per time step.
  2. [Section IV-B, Fig. 6] The caption and text state that 'an obvious outlier result was removed' for libfabric at 8192 nodes, but do not report the value of the removed measurement or the criterion for calling it an outlier; please provide this information for reproducibility.
  3. [Section VI] There is a typo in 'more sophisticated subnet-architecutres'; it should read 'architectures'.
  4. [Section IV-D, reference [66]] The discussion of PyTorch DDP scaling would be stronger if it cited the specific version of PyTorch used and clarified whether the socket limit was observed with the ROCm build; the current reference [66] is about large language models and may not be the most direct support.
  5. [Section V-B] The sentence 'The agreement is good enough to unambiguously classify the region of origin...' is a strong claim that goes beyond the qualitative plot; please either provide a quantitative classifier evaluation or soften this statement.

Circularity Check

1 steps flagged · score 4.0 of 10

The streaming-workflow result is independently measured, but the scientific 'prediction' of particle dynamics from radiation is evaluated on the same in-transit stream used for training, with no stated held-out split.

  1. fitted input called prediction [Section IV-C (training buffer / experience replay) and Section V-B (Quantifying the predictive capabilities)]
    "While training on the continuous data stream of non-steady configurations, we employ experience replay (EP) [58] to avoid catastrophic forgetting of earlier simulation time steps while training on later ones. ... To evaluate the performance of the trained model, we invert the radiation spectra back to the original momentum distribution, focusing on the momentum component px in the following discussion."

    The network is trained on the simulation stream itself, with the training buffer deliberately retaining earlier time steps via experience replay, and all training runs are performed on the same PIConGPU stream that later supplies the 'selected example sub-volume' for evaluation. The paper does not specify any held-out split or external test set. Consequently, the reported inversion agreement in Fig. 9 measures how well the model reproduces the distribution it was fitted on, rather than an independent predictive capability. This is not an equation-level tautology, but the fitted network is being presented as a 'prediction' on its own training distribution.

full rationale

The central engineering claim — that a PIConGPU simulation can stream data in-transit via openPMD/ADIOS2 to a PyTorch ML application without a filesystem bottleneck — is supported by direct measurements: full-system no-op streaming reaches 20-30 TB/s (Section IV-B), and the actual PIConGPU+MLapp pipeline is trained and timed on 8-96 nodes (Section V-A). Those numbers are independent of the ML model's inversion accuracy, so the workflow feasibility result is not circular. The disclosed N/RCCL socket limit beyond 100 nodes is a scaling limitation, not a circularity. The only partial circularity is in the scientific proof-of-concept: the model is trained and evaluated on the same in-transit simulation stream, with no held-out split, so the claimed 'learning correlations' / 'predictive capabilities' partly reduce to training-set fit. The self-citation to [43] for equating no-op throughput with real throughput is load-bearing for the full-system extrapolation, but the underlying benchmark is a direct measurement and the actual training pipeline has its own smaller-scale measurements; this is better treated as a correctness/coverage concern than as definitional circularity.

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

The central demonstration rests on several modeling and measurement assumptions: the correctness of the far-field radiation plugin, the representativeness of the sampled particles, the transferability of the no-op throughput benchmark, and the representativeness of the single evaluation sub-volume. Loss weights and training hyperparameters are hand-tuned. No new physical entities are introduced; the latent representation is a model-internal construct, not an independently evidenced entity.

free parameters (5)
  • Loss weights in Eq. (1) = 1.0, 0.001, 0.3, 40, 0.03
    Stated as empirically tuned in Section IV-C; these weights control the balance between reconstruction, KL, MSE, and MMD terms and affect all reported model behavior.
  • Learning rates l_VAE and l_INN, base learning rate l_base = l_base = 1e-6, l_VAE = m_VAE * l_INN
    Section V-A.1 reports that separate rates were needed at large scale and that hyperparameters do not transfer from small-scale runs.
  • Experience replay buffer sizes and sampling counts = N_now=10, N_EP=20, n_now=4, n_EP=4, nrep up to 96
    Section IV-C: buffer sizes and replay schedule are design choices that determine continual-learning behavior; no ablation is shown.
  • Architecture sizes = latent 544, features 6->16->32->64->128->256->608, decoder 1024->4096 particles, INN 272->256->544
    Section IV-C: these dimensions define model capacity; no systematic architecture search is reported.
  • Outlier removal thresholds = 4 sigma; one 'obvious outlier' removed
    Section V-A and Fig. 6: reported scaling efficiency and throughput use filtered measurements.
assumptions (5)
  • domain assumption The PIConGPU far-field radiation plugin computes observationally correct spectra via the Lienard-Wiechert potential approach.
    Section IV-A uses this plugin as ground-truth radiation input for the ML model; its correctness is assumed from prior radiation-diagnostics work.
  • domain assumption The PIConGPU simulation of the relativistic Kelvin-Helmholtz instability provides sufficient ground truth for both particle dynamics and radiation.
    Section II chooses KHI because its dynamics are well understood, and all training targets and validation plots come from the simulation, not from experiments.
  • ad hoc to paper The no-op consumer full-system benchmark approximates the real MLapp streaming throughput.
    Section IV-B states the no-op consumer performs no computation besides measuring performance; the full-system throughput claim depends on this proxy.
  • ad hoc to paper A 3e4-particle sample is a sufficient representation of local phase-space dynamics for the autoencoder and inversion.
    Section IV-C feeds fixed 3e4-particle vectors and notes that random sampling of fewer particles would not suffice; no sampling-error analysis is given.
  • ad hoc to paper The selected evaluation sub-volume in Fig. 9 is representative of model performance across the simulation.
    Section V-B evaluates one selected example sub-volume without stating a holdout protocol, repeated runs, or uncertainty quantification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Artificial Scientist -- in-transit Machine Learning of Plasma Simulations." pith.science (2026). https://pith.science/paper/DLNR4TN7

@misc{pith2026250103383,
  author       = {Pith},
  title        = {Pith review of: The Artificial Scientist -- in-transit Machine Learning of Plasma Simulations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DLNR4TN7}},
  note         = {Machine review of arXiv:2501.03383}
}
read the original abstract

Increasing HPC cluster sizes and large-scale simulations that produce petabytes of data per run, create massive IO and storage challenges for analysis. Deep learning-based techniques, in particular, make use of these amounts of domain data to extract patterns that help build scientific understanding. Here, we demonstrate a streaming workflow in which simulation data is streamed directly to a machine-learning (ML) framework, circumventing the file system bottleneck. Data is transformed in transit, asynchronously to the simulation and the training of the model. With the presented workflow, data operations can be performed in common and easy-to-use programming languages, freeing the application user from adapting the application output routines. As a proof-of-concept we consider a GPU accelerated particle-in-cell (PIConGPU) simulation of the Kelvin- Helmholtz instability (KHI). We employ experience replay to avoid catastrophic forgetting in learning from this non-steady process in a continual manner. We detail challenges addressed while porting and scaling to Frontier exascale system.

Figures

Figures reproduced from arXiv: 2501.03383 by the authors.

Figure 1
Figure 1. 3D rendering of the KHI simulated with PIConGPU [6] and rendered with ISAAC [7]. The middle graphic shows electrons as particles that radiate strongly, with the radiation intensity indicated from blue to red. The initial plasma flow direction is depicted as arrows, blue marking propagation towards and red away from the radiation detector. The spectrally resolved radiation determined by the synthetic radiation detect… view at source ↗
Figure 2
Figure 2. Tasks that an ML model can be trained on, based on the simulation data in-transit: (a) solving the inverse problem of predicting particle dynamics from radiation signatures; (b) extracting features of particle dynamics and reconstruction thereof (represen￾tation learning/compression); (c) surrogate model of radiation emit￾ted by complex particle dynamics. Rounded boxes ( ) indicate inputs/outputs, sharp boxes ( ) re… view at source ↗
Figure 3
Figure 3. Three aspects to streaming between loosely coupled producer and consumer: (a) streaming without going through storage unlocks more bandwidth; (b) reducing simulation data close to the producer lowers bandwidth requirements; (c) to distributed producer and con￾sumer, system topology presents communication paths with vastly different bandwidths which must be reconciled with the loosely￾coupled application’s communicat… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: FOM scaling of PIConGPU from 24 GPUs (6 nodes) to 36 864 GPUs (9216 nodes) on Frontier (red line) using a general test case [30]. PIConGPU achieves an average FOM for the largest run of 65.3 TeraUpdates/s vs 14.7 TeraUpdates/s on Summit. for the first time introduces a…
Figure 5
Figure 5. Figure 5: High-level overview of the software stack for data exchange between PIConGPU and the MLapp. ADIOS2, all supported by the reference implementation openPMD-api, of which we further consider ADIOS2 in this study due to its support for in-transit I/O in HPC settings, as sh…
Figure 7
Figure 7. Figure 7: Block-wise architecture of the ML model encompassing all three tasks summarized in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Frontier weak scaling of the in-transit training from 32 to 384 GCDs (8 to 96 nodes). Measurements are single-batch times averaged over multiple runs and over iterations within runs after removal of > 4σ outliers. sufficient examples available in the dataset. When the …
Figure 9
Figure 9. Figure 9: Comparing the ground truth data from PIConGPU with the prediction of the ML model on a selected example sub-volume. Blue & red are undisturbed plasma streams approaching or receding from the detector, green depicts the KHI vortex regions. (a) radiation spec￾tra, (b) gr…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Exascale Implicit Kinetic Plasma Simulations on El~Capitan for Solving the Micro-Macro Coupling in Magnetospheric Physics

    cs.CE 2025-07 conditional novelty 5.0 of 10

    iPIC3D, an implicit kinetic plasma code, scales to 32,768 AMD MI300A APUs and projects 22.4 PFLOP/s sustained, reaching simulation domains of hundreds of ion skin depths.

Reference graph

Works this paper leans on

69 extracted references · 64 canonical work pages · cited by 1 Pith paper

  1. [1]

    AI-Augmented Facilities: Bridging Experiment and Simulation with ML (Dagstuhl Seminar 23132),

    P.-T. Bremer, B. Spears, T. Gibbs, and M. Bussmann, “AI-Augmented Facilities: Bridging Experiment and Simulation with ML (Dagstuhl Seminar 23132),” Dagstuhl Reports, vol. 13, no. 3, pp. 106–131, 2023

    work page 2023

  2. [2]

    Deep learning for digital holography: a review,

    T. Zeng, Y . Zhu, and E. Y . Lam, “Deep learning for digital holography: a review,” Opt. Express, vol. 29, pp. 40572–40593, Nov 2021

    work page 2021

  3. [3]

    Neural Network Potentials: A Concise Overview of Methods,

    E. Kocer, T. W. Ko, and J. Behler, “Neural Network Potentials: A Concise Overview of Methods,” Annual Review of Physical Chemistry , vol. 73, pp. 163–186, 2022

    work page 2022

  4. [4]

    Data-driven Science and Machine Learning Methods in Laser-Plasma Physics,

    A. D ¨opp, C. Eberle, S. Howard, F. Irshad, J. Lin, and M. Streeter, “Data-driven Science and Machine Learning Methods in Laser-Plasma Physics,” High Power Laser Science and Engineering , pp. 1–50, 2023

    work page 2023

  5. [5]

    Machine learning for anomaly detection in particle physics,

    V . Belis, P. Odagiu, and T. K. Aarrestad, “Machine learning for anomaly detection in particle physics,” Reviews in Physics , vol. 12, p. 100091, 2024

    work page 2024

  6. [6]

    Radiative signatures of the relativistic Kelvin-Helmholtz instability,

    M. Bussmann, H. Burau, T. E. Cowan, A. Debus, A. Huebl, G. Juck- eland, T. Kluge, W. E. Nagel, R. Pausch, F. Schmitt, U. Schramm, J. Schuchart, and R. Widera, “Radiative signatures of the relativistic Kelvin-Helmholtz instability,” in SC ’13 Proc. International Conference for High Performance Computing, Networking, Storage and Analysis , pp. 5–1 – 5–12, 2013

    work page 2013

  7. [7]

    Hardware-agnostic interactive ex- ascale in situ visualization of particle-in-cell simulations,

    F. Meyer, B. Hernandez, R. Pausch, R. Widera, D. Groß, S. Bas- trakov, A. Huebl, G. Juckeland, J. Kelling, M. Leinhauser, D. Rogers, U. Schramm, K. Steiniger, S. Gumhold, J. Young, M. Bussmann, S. Chandrasekaran, and A. Debus, “Hardware-agnostic interactive ex- ascale in situ visualization of particle-in-cell simulations,” in Proc. Platform for Advanced S...

    work page 2023

  8. [8]

    Stability of Hydromagnetic Kelvin-Helmholtz Discontinu- ity,

    A. K. Sen, “Stability of Hydromagnetic Kelvin-Helmholtz Discontinu- ity,” Physics of Fluids , vol. 6, no. 8, p. 1154, 1963

    work page 1963

Show all 69 references
  1. [9]

    Goodfellow, Y

    I. Goodfellow, Y . Bengio, and A. Courville, Deep Learning. MIT Press,

  2. [10]

    Machine learning for streaming data: state of the art, challenges, and opportunities,

    H. M. Gomes, J. Read, A. Bifet, J. P. Barddal, and J. a. Gama, “Machine learning for streaming data: state of the art, challenges, and opportunities,” SIGKDD Explor. Newsl., vol. 21, p. 6–22, nov 2019

    2019

  3. [11]

    A Comprehensive Survey of Continual Learning: Theory, Method and Application,

    L. Wang, X. Zhang, H. Su, and J. Zhu, “A Comprehensive Survey of Continual Learning: Theory, Method and Application,” 2024

    2024

  4. [12]

    On the scalability of data reduction techniques in current and upcoming hpc systems from an application perspective,

    A. Huebl, R. Widera, F. Schmitt, A. Matthes, N. Podhorszki, J. Y . Choi, S. Klasky, and M. Bussmann, “On the scalability of data reduction techniques in current and upcoming hpc systems from an application perspective,” in High Performance Computing (J. M. Kunkel, R. Yokota, M...

    2017

  5. [13]

    Improving i/o performance for exascale appli- cations through online data layout reorganization,

    L. Wan, A. Huebl, J. Gu, F. Poeschel, A. Gainaru, R. Wang, J. Chen, X. Liang, D. Ganyushin, T. Munson, I. Foster, J.-L. Vay, N. Podhorszki, K. Wu, and S. Klasky, “Improving i/o performance for exascale appli- cations through online data layout reorganization,” IEEE Trans. Para...

    2022

  6. [14]

    Big data multimedia mining: feature extraction facing volume, velocity, and variety,

    V . Pandit, S. Amiriparian, M. Schmitt, A. Mousa, and B. Schuller, “Big data multimedia mining: feature extraction facing volume, velocity, and variety,”Big Data Analytics for Large-Scale Multimedia Search, vol. 61, 2019

    2019

  7. [15]

    Transitioning from file-based hpc workflows to streaming data pipelines with openpmd and adios2,

    F. Poeschel, J. E, W. F. Godoy, N. Podhorszki, S. Klasky, G. Eisenhauer, P. E. Davis, L. Wan, A. Gainaru, J. Gu, F. Koller, R. Widera, M. Buss- mann, and A. Huebl, “Transitioning from file-based hpc workflows to streaming data pipelines with openpmd and adios2,” in Driving Sci...

    2022

  8. [16]

    Frontier

    OLCF, “Frontier.” https://www.olcf.ornl.gov/frontier/, 2024

    2024

  9. [17]

    PUNCH4NFDI Use case class 5: real-time challenges, data irreversibil- ity,

    “PUNCH4NFDI Use case class 5: real-time challenges, data irreversibil- ity,” 2024

    2024

  10. [18]

    PyTorch Distributed Data Parallel Documentaion

    “PyTorch Distributed Data Parallel Documentaion.” https://pytorch.org/ docs/2.0/notes/ddp.html

  11. [19]

    Kelvin-helmholtz instability at saturn’s magnetopause: Hybrid simulations,

    P. Delamere, R. Wilson, and A. Masters, “Kelvin-helmholtz instability at saturn’s magnetopause: Hybrid simulations,” Journal of Geophysical Research: Space Physics , vol. 116, no. A10, 2011

    2011

  12. [20]

    Recent progress in quantifying hydrodynamics instabil- ities and turbulence in inertial confinement fusion and high-energy- density experiments,

    A. Casner, “Recent progress in quantifying hydrodynamics instabil- ities and turbulence in inertial confinement fusion and high-energy- density experiments,” Philosophical Transactions of the Royal Society A, vol. 379, no. 2189, p. 20200021, 2021

    2021

  13. [21]

    Faster ablative kelvin–helmholtz instability growth in a magnetic field,

    J. D. Sadler, S. Green, S. Li, Y . Zhou, K. A. Flippo, and H. Li, “Faster ablative kelvin–helmholtz instability growth in a magnetic field,”Physics of Plasmas, vol. 29, no. 5, 2022

    2022

  14. [22]

    dc-magnetic- field generation in unmagnetized shear flows,

    T. Grismayer, E. P. Alves, R. a. Fonseca, and L. O. Silva, “dc-magnetic- field generation in unmagnetized shear flows,” Physical Review Letters, vol. 111, p. 015005, 7 2013

    2013

  15. [23]

    Large-Scale Magnetic Field Generation Via the Kinetic Kelvin-Helmholtz Instability in Unmagnetized Scenarios,

    E. P. Alves, T. Grismayer, S. F. Martins, F. Fi ´uza, R. a. Fonseca, and L. O. Silva, “Large-Scale Magnetic Field Generation Via the Kinetic Kelvin-Helmholtz Instability in Unmagnetized Scenarios,” The Astrophysical Journal Letters, vol. 746, no. 2, p. L14, 2012

    2012

  16. [24]

    Identifying the linear phase of the relativistic kelvin-helmholtz instability and measuring its growth rate via radiation,

    R. Pausch, M. Bussmann, A. Huebl, U. Schramm, K. Steiniger, R. Widera, and A. Debus, “Identifying the linear phase of the relativistic kelvin-helmholtz instability and measuring its growth rate via radiation,” Physical Review E , vol. 96, p. 013316, 7 2017

    2017

  17. [25]

    PICon- GPU: A fully relativistic particle-in-cell code for a GPU cluster,

    H. Burau, R. Widera, W. H ¨onig, G. Juckeland, A. Debus, T. Kluge, U. Schramm, T. E. Cowan, R. Sauerbrey, and M. Bussmann, “PICon- GPU: A fully relativistic particle-in-cell code for a GPU cluster,” IEEE Trans. Plasma Sci. , vol. 38, no. 10 PART 2, pp. 2831–2839, 2010

    2010

  18. [26]

    Alpaka – an abstraction library for parallel kernel acceleration,

    E. Zenker, B. Worpitz, R. Widera, A. Huebl, G. Juckeland, A. Kn ¨upfer, W. E. Nagel, and M. Bussmann, “Alpaka – an abstraction library for parallel kernel acceleration,” in IEEE Intl. Parallel and Distributed Processing Symposium Workshops (IPDPSW) , pp. 631–640, 5 2016

    2016

  19. [27]

    R. W. Hockney and J. W. Eastwood, Computer simulation using parti- cles. Bristol and New York: Adam Hilger - IOP Publishing Ltd, 1988

    1988

  20. [28]

    S. I. Kabanikhin, Inverse and ill-posed problems: theory and applica- tions. de Gruyter, 2011

    2011

  21. [29]

    Frontier User Guide: System Overview

    “Frontier User Guide: System Overview.” https://docs.olcf.ornl.gov/ systems/frontier user guide.html#frontier-compute-node

  22. [30]

    TWEAC test-case PIConGPU

    “TWEAC test-case PIConGPU.” https://github.com/ ComputationalRadiationPhysics/picongpu/tree/dev/share/picongpu/ benchmarks/TWEAC-FOM, 2023

    2023

  23. [31]

    Ready for the frontier: Preparing applications for the world’s first exascale system,

    R. D. Budiardja, M. Berrill, M. Eisenbach, G. R. Jansen, W. Joubert, S. Nichols, D. M. Rogers, A. Tharrington, and O. Bronson Messer, “Ready for the frontier: Preparing applications for the world’s first exascale system,” in International Conference on High Performance Computi...

    2023

  24. [32]

    A Generic Approach for Developing Highly Scalable Particle-Mesh Codes for GPUs,

    W. H ¨onig, F. Schmitt, R. Widera, H. Burau, G. Juckeland, M. S. M¨uller, and M. Bussmann, “A Generic Approach for Developing Highly Scalable Particle-Mesh Codes for GPUs,” SAAHPC.–2010, 2010

    2010

  25. [33]

    libPMacc

    “libPMacc.” https://github.com/ComputationalRadiationPhysics/ picongpu/tree/dev/include/pmacc, 2023

    2023

  26. [34]

    openpmd-api: C++ & python api for scientific i/o with openpmd,

    F. Koller, F. Poeschel, J. Gu, and A. Huebl, “openpmd-api: C++ & python api for scientific i/o with openpmd,” 2018

    2018

  27. [35]

    openPMD 1.1.0: Base paths for mesh- and particle- only files and updated attributes,

    A. Huebl, R. Lehe, J.-L. Vay, D. P. Grote, I. Sbalzarini, S. Kuschel, D. Sagan, F. P´erez, F. Koller, and M. Bussmann, “openPMD 1.1.0: Base paths for mesh- and particle- only files and updated attributes,” Feb. 2018

    2018

  28. [36]

    Adios 2: The adaptable input output system. a framework for high- performance data management,

    W. F. Godoy, N. Podhorszki, R. Wang, C. Atkins, G. Eisenhauer, J. Gu, P. Davis, J. Choi, K. Germaschewski, K. Huck, A. Huebl, M. Kim, J. Kress, T. Kurc, Q. Liu, J. Logan, K. Mehta, G. Ostrouchov, M. Parashar, F. Poeschel, D. Pugmire, E. Suchyta, K. Takahashi, N. Thompson, S. T...

    2020

  29. [37]

    How to test and verify radiation diagnostics simulations within particle-in-cell frameworks,

    R. Pausch, A. Debus, R. Widera, K. Steiniger, A. Huebl, H. Burau, M. Bussmann, and U. Schramm, “How to test and verify radiation diagnostics simulations within particle-in-cell frameworks,” Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectromet...

    2014

  30. [38]

    J. D. Jackson, Classical Electrodynamics. New York: John Wiley and Sons, Inc., third ed., 1998

    1998

  31. [39]

    Quantitatively consistent computation of coherent and incoherent radiation in particle-in-cell codes - a general form factor formalism for macro-particles,

    R. Pausch, A. Debus, A. Huebl, U. Schramm, K. Steiniger, R. Widera, and M. Bussmann, “Quantitatively consistent computation of coherent and incoherent radiation in particle-in-cell codes - a general form factor formalism for macro-particles,” Nuclear Instruments and Methods in...

    2018

  32. [40]

    Computing Angularly-Resolved Far-Field Emission Spectra In Particle-In-Cell Codes Using GPUs,

    R. Pausch, H. Burau, M. Bussmann, J. Couperus, T. E. Cowan, A. De- bus, A. Huebl, A. Irman, A. K ¨ohler, U. Schramm, K. Steiniger, and R. Widera, “Computing Angularly-Resolved Far-Field Emission Spectra In Particle-In-Cell Codes Using GPUs,” in Proc. IPAC2014, pp. 761– 764, 2014

    2014

  33. [41]

    Overview on projects around openPMD

    “Overview on projects around openPMD.” https://github.com/openPMD/ openPMD-projects, 2024

    2024

  34. [42]

    Understanding the impact of data staging for coupled scientific workflows,

    A. Gainaru, L. Wan, R. Wang, E. Suchyta, J. Chen, N. Podhorszki, J. Kress, D. Pugmire, and S. Klasky, “Understanding the impact of data staging for coupled scientific workflows,” IEEE Trans. Parallel Distrib. Syst., vol. 33, no. 12, pp. 4134–4147, 2022

    2022

  35. [43]

    Stream- ing data in hpc workflows using adios,

    G. Eisenhauer, N. Podhorszki, A. Gainaru, S. Klasky, P. E. Davis, M. Parashar, M. Wolf, E. Suchtya, E. Fredj, V . Bolea, F. P ¨oschel, K. Steiniger, M. Bussmann, R. Pausch, and S. Chandrasekaran, “Stream- ing data in hpc workflows using adios,” 2024

    2024

  36. [44]

    Open Fabric Interfaces

    “Open Fabric Interfaces.” https://github.com/ofiwg/libfabric/

  37. [45]

    OLCF announces storage specifications for frontier exascale system

    “OLCF announces storage specifications for frontier exascale system.” https://www.olcf.ornl.gov/2021/05/20/olcf-announces-storage- specifications-for-frontier-exascale-system/, 2021

    2021

  38. [46]

    Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning,

    Y . Gal and Z. Ghahramani, “Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning,” 2016

    2016

  39. [47]

    Density estimation using Real NVP,

    L. Dinh, J. Sohl-Dickstein, and S. Bengio, “Density estimation using Real NVP,” CoRR, vol. abs/1605.08803, 2016

    2016 arXiv

  40. [48]

    Structured Output Learning with Conditional Generative Flows,

    Y . Lu and B. Huang, “Structured Output Learning with Conditional Generative Flows,” 2020

    2020

  41. [49]

    Analyzing Inverse Problems with Invertible Neural Networks,

    L. Ardizzone, J. Kruse, S. J. Wirkert, D. Rahner, E. W. Pellegrini, R. S. Klessen, L. Maier-Hein, C. Rother, and U. K ¨othe, “Analyzing Inverse Problems with Invertible Neural Networks,” CoRR, vol. abs/1808.04730, 2018

    2018 arXiv

  42. [50]

    Density estimation by dual ascent of the log-likelihood,

    E. Tabak and E. Vanden-Eijnden, “Density estimation by dual ascent of the log-likelihood,” Communications in Mathematical Sciences , vol. 8, no. 1, pp. 217–233, 2010

    2010

  43. [51]

    A Family of Nonparametric Density Estimation Algorithms,

    E. G. Tabak and C. V . Turner, “A Family of Nonparametric Density Estimation Algorithms,” Communications on Pure and Applied Mathe- matics, vol. 66, no. 2, pp. 145–164, 2013

    2013

  44. [52]

    NICE: Non-linear Independent Components Estimation,

    L. Dinh, D. Krueger, and Y . Bengio, “NICE: Non-linear Independent Components Estimation,” 2015

    2015

  45. [53]

    Generative Models for 3D Point Clouds,

    L. Kong, P. Rajak, and S. Shakeri, “Generative Models for 3D Point Clouds,” 2023

    2023

  46. [54]

    PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation,” 2017

    2017

  47. [55]

    Glow: Generative Flow with Invertible 1x1 Convolutions,

    D. P. Kingma and P. Dhariwal, “Glow: Generative Flow with Invertible 1x1 Convolutions,” in Advances in Neural Information Processing Systems (S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa- Bianchi, and R. Garnett, eds.), vol. 31, Curran Associates, Inc., 2018

    2018

  48. [56]

    A Point Set Generation Network for 3D Object Reconstruction from a Single Image,

    H. Fan, H. Su, and L. Guibas, “A Point Set Generation Network for 3D Object Reconstruction from a Single Image,” 2016

    2016

  49. [57]

    Auto-Encoding Variational Bayes,

    D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,” 2013

    2013

  50. [58]

    On tiny episodic memories in continual learning,

    A. Chaudhry, M. Rohrbach, M. Elhoseiny, T. Ajanthan, P. K. Dokania, P. H. S. Torr, and M. Ranzato, “On tiny episodic memories in continual learning,” arXiv: Learning, 2019

    2019

  51. [59]

    Adam: A Method for Stochastic Optimization,

    D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,” 2017

    2017

  52. [60]

    One weird trick for parallelizing convolutional neural networks,

    A. Krizhevsky, “One weird trick for parallelizing convolutional neural networks,” 2014

    2014

  53. [61]

    Learning Rep- resentations and Generative Models for 3D Point Clouds,

    P. Achlioptas, O. Diamanti, I. Mitliagkas, and L. Guibas, “Learning Rep- resentations and Generative Models for 3D Point Clouds,” in Proc. 35th International Conference on Machine Learning (J. Dy and A. Krause, eds.), vol. 80 of Proc. Machine Learning Research , pp. 40–49, PML...

    2018

  54. [62]

    The Earth Mover’s Distance as a Metric for Image Retrieval,

    Y . Rubner, C. Tomasi, and L. J. Guibas, “The Earth Mover’s Distance as a Metric for Image Retrieval,” International Journal of Computer Vision, vol. 40, pp. 99–121, Nov 2000

    2000

  55. [63]

    Interpolating between Optimal Transport and MMD using Sinkhorn Divergences,

    J. Feydy, T. S ´ejourn´e, F.-X. Vialard, S.-i. Amari, A. Trouve, and G. Peyr ´e, “Interpolating between Optimal Transport and MMD using Sinkhorn Divergences,” in The 22nd International Conference on Arti- ficial Intelligence and Statistics , pp. 2681–2690, 2019

    2019

  56. [64]

    Kernel Operations on the GPU, with Autodiff, without Memory Overflows,

    B. Charlier, J. Feydy, J. A. Glaun `es, F.-D. Collin, and G. Durif, “Kernel Operations on the GPU, with Autodiff, without Memory Overflows,” Journal of Machine Learning Research , vol. 22, no. 74, pp. 1–6, 2021

    2021

  57. [65]

    Fast geometric learning with symbolic matrices,

    J. Feydy, J. Glaun `es, B. Charlier, and M. Bronstein, “Fast geometric learning with symbolic matrices,” Advances in Neural Information Processing Systems, vol. 33, 2020

    2020

  58. [66]

    Optimizing distributed training on frontier for large language models,

    S. Dash, I. Lyngaas, J. Yin, X. Wang, R. Egele, G. Cong, F. Wang, and P. Balaprakash, “Optimizing distributed training on frontier for large language models,” 2023

    2023

  59. [67]

    Learning representations and generative models for 3D point clouds,

    P. Achlioptas, O. Diamanti, I. Mitliagkas, and L. Guibas, “Learning representations and generative models for 3D point clouds,” inProc. 35th International Conference on Machine Learning (J. Dy and A. Krause, eds.), vol. 80 of Proc. Machine Learning Research , pp. 40–49, PMLR, ...

    2018

  60. [68]

    Emerging Properties in Self-Supervised Vision Transform- ers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging Properties in Self-Supervised Vision Transform- ers,” 2021

    2021

  61. [2016]

    http://www.deeplearningbook.org

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.