Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Predictive Hydrodynamic Simulations for Laser Direct-drive Implosion Experiments via Artificial Intelligence

T0 review · 4 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 single 65% absorption factor lets an AI surrogate predict DCI implosion dynamics, matching collision time to about 0.1 ns.

desk verdict A useful engineering proof-of-concept for AI-surrogate calibration in direct-drive ICF, but the headline predictive claim rests on a constant calibrated absorption factor and a single holdout shot. read the letter →

arxiv 2507.16227 v1 pith:U6PO72IP submitted 2025-07-22 physics.plasm-ph cs.AI

Zixu Wang , Yuhan Wang , Junfei Ma , Fuyuan Wu , Junchi Yan , Xiaohui Yuan , Zhe Zhang , Jie Zhang This is my paper
classification physics.plasm-phcs.AI PACS 52.57.-z
keywords double-coneignitionlaserdirect-driveimplosiondynamicspredictivesimulationTransformersurrogatephysics-informedsamplingeffectiveabsorption
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

The paper aims to show that a fast neural-network surrogate can replace slow one-dimensional hydrodynamic simulations for laser direct-drive implosion experiments, and that a single physics-motivated calibration constant—an effective laser absorption rate of about 65%—suffices to make the simulations match measured implosion dynamics. The authors build MULTI-Net, a Transformer-based model that maps laser waveforms and target radius to implosion features such as collision time, mean velocity, and areal density. They also introduce a Physics-Informed Decoder (PID) sampling method that generates training data concentrated in physically relevant regions, cutting prediction error by roughly 82% on average compared to Latin hypercube sampling. The payoff is demonstrated on double-cone ignition shots: after calibrating on three shots, the model predicts a fourth shot's collision time to within about 0.1 ns and yields a mean implosion velocity of 195 km/s and a peak collision density of 117 g/cc, consistent with X-ray streak camera measurements.

What carries the argument

The load-bearing machinery combines three components. MULTI-Net is an encoder-only Transformer whose multi-head self-attention captures long-range dependencies in the 100-point laser power waveform; attention pooling aggregates the sequence to a fixed-size representation and a two-layer MLP predicts the implosion features. The Physics-Informed Decoder (PID) is a small network trained to reconstruct laser waveforms from implosion features, then used to generate training samples that are nearly uniform in the physically meaningful feature space, avoiding the high-dimensional exponential explosion that plain Latin hypercube sampling suffers. The third component is a calibration step that multiplies the experimental laser waveform by a constant absorption factor $\eta$ and minimizes a weighted squared error between simulated and measured implosion features (here, collision time) over three calibration shots. The Transformer architecture improves the median residual by 88.2% on average relative to a same-size MLP, and using PID data reduces validation error by 82.4% on average.

What would settle it

Run a DCI-R10 shot with a laser energy or pulse shape substantially different from the three calibration shots, extract the experimental collision time from the X-ray streak camera, and check whether the calibrated simulation with a fixed 65% absorption still matches within about 0.1 ns; a systematic deviation would invalidate the constant-absorption assumption.

Watch

Extended reading notes

Core claim

The central claim is that the gap between one-dimensional hydrodynamic simulations and real double-cone ignition experiments can be closed by a single global scaling factor applied to the laser energy. Setting the effective absorption rate to about 65% (64.5% for the MULTI-IFE simulator, 65.5% for the MULTI-Net surrogate) makes the simulated collision time match three calibration shots, and the same factor transfers to a fourth shot with different target and laser parameters. With this calibration, the simulation reproduces the measured collision time (6.17 ns experimental versus 6.11 ns predicted), the plasma confinement duration of about 0.4 ns, and the shell acceleration behavior seen in X-ray emission between 2 and 4 ns. The authors infer from the calibrated simulations that the mean implosion velocity is about 195 km/s, the peak collision density is about 117 g/cc, and the peak areal density is about 0.48 g/cm².

Load-bearing premise

The prediction for shot 33 rests on treating the effective laser absorption rate of about 65% as a constant that does not depend on laser power or shot-to-shot variations, so the value fitted on three shots is simply carried over to the target being predicted.

Editorial extensions

If this is right

  • The MULTI-Net surrogate computes implosion features almost instantly, making exhaustive calibration scans and rapid shot-design iteration practical instead of running hours of hydrodynamics per case.
  • Because the model takes a laser waveform plus target radius as input, the same architecture can be retrained for other implosion campaigns or other 1D code datasets, not just DCI-R10.
  • The calibrated simulation supplies experimentally difficult-to-measure quantities—mean implosion velocity, peak collision density, and areal density—that characterize the collision state of DCI targets.
  • The inferred ~65% absorption factor gives a quantitative energy-coupling budget for the DCI-R10 campaign, attributing roughly 35% of delivered laser energy to losses from geometry and laser-plasma instabilities.
  • The agreement on shot 33 supports using the calibrated simulation-experiment cycle to pre-test future laser waveforms and target parameters before firing.

Reading between the lines

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

  • If the real laser-energy coupling varies with shot conditions, the constant 65% absorption will need to become a shot-dependent parameter; one way to test this is to calibrate on shots with deliberately different pulse shapes and see whether a single factor still collapses the collision-time error.
  • The PID sampling idea—sample in output space, decode back to input space—is a general recipe for any expensive simulation with high-dimensional inputs and a few physically meaningful outputs, so it could be reused beyond laser fusion.
  • Once two-dimensional simulations of the DCI collision become available, the same Transformer-plus-PID pipeline could be retrained on spatial density and temperature fields, extending predictions from integral features to full plasma profiles.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes an AI-empowered workflow for predicting laser direct-drive implosion dynamics, using the double-cone ignition (DCI) scheme as a test case. A Transformer-based surrogate model, MULTI-Net, is trained on 20,000 one-dimensional MULTI-IFE simulations to predict implosion features (mean velocity, areal density, in-flight aspect ratio, maximum density, collision time) from laser waveforms and target thickness. A Physics-Informed Decoder (PID) sampling method is introduced to improve dataset quality, and the surrogate is used to accelerate calibration of an effective laser absorption factor η. Calibration on three DCI-R10 shots (22, 24, 37) yields η ≈ 65%, and the calibrated simulations are then applied to a fourth shot (33). For shot 33, the measured collision time of 6.17 ns is matched to within about 0.1 ns, and the calibrated simulation gives a mean implosion velocity of 195 km/s and a peak collision density of 117 g/cc. The paper concludes that the framework enhances the predictive ability of simulations for DCI experiments.

Significance. If the central claim is robust, the workflow would be a valuable contribution to inertial-confinement-fusion research, particularly for cases where high-dimensional parameter spaces make exhaustive simulation impractical. The paper has several concrete strengths: the Transformer surrogate achieves high regression accuracy (R² ≈ 0.9993 for Vmean and ≈ 0.9966 for tcol in the direct comparison of Figure 3); the PID sampling method is an interesting and plausible idea, with clear evidence that it reduces prediction error on a common-size validation set by 82.4% on average; and the use of actual experimental x-ray streak data on the SG-II Upgrade facility anchors the study in a real application. The key weakness is that the experimental validation is thin: the calibration relies on three shots, the transfer to a fourth shot is tested exactly once, the inferred velocity and density are not independently measured, and no uncertainty quantification is provided. The central claim is therefore plausible but not yet fully supported.

major comments (4)
  1. [Section 5.1, Figure 6] The calibration of the effective absorption rate η from only three shots (22, 24, 37), with the assumption that η is a constant independent of laser power, is load-bearing for the entire prediction. The paper does not provide a leave-one-out cross-validation over the four DCI-R10 shots, nor does it test whether η transfers to shot 33 in a statistically meaningful way. Given that the x-ray streak camera temporal resolution is 130 ps (Section 5.2), the post-calibration MAE of about 0.1 ns is comparable to the instrumental noise floor, so the three calibration shots cannot resolve shot-to-shot variations in η smaller than that noise. A single held-out shot with 0.06–0.1 ns agreement could be coincidental. I recommend adding a leave-one-out analysis over the four R10 shots and a sensitivity study showing how the predicted collision time and velocity/density change when η is varied by ±5%; without this, the claim that a constant 65% absorption is 'suitable for the DCI-R10 experiments' is not adequately supported.
  2. [Section 5.2, Figure 7] The headline numbers 195 km/s and 117 g/cc are not experimental measurements; they are outputs of the calibrated simulation. The x-ray streak camera provides the trajectory and timing of the CD plasma, from which the collision time and confinement duration are inferred, but the shell velocity and collided plasma density are quantities that the authors 'infer from the simulations' (their own wording). No independent diagnostic—such as Doppler velocimetry, time-resolved radiography, or spectroscopy—is presented to validate these values. The abstract and conclusions present 195 km/s and 117 g/cc as predicted results without clearly distinguishing simulated inference from direct measurement. This distinction should be made explicit in the abstract, and the conclusion should be tempered to state that these values are simulation-inferred estimates pending independent confirmation.
  3. [Section 3.2 vs Abstract and Section 5.2] The abstract and introduction state that MULTI-Net predicts implosion features according to 'laser waveforms and target radius', but Section 3.2 specifies that the training dataset uses a fixed outer radius of 550 µm and that the model input consists of 100 laser power points plus one target layer thickness—no radius input is listed. Shot 33, however, has an outer radius of 560 µm, which is outside the training configuration (and also outside the 50–110 µm thickness range? It is within, 60 µm, but radius is outside). The paper reports MULTI-Net predictions for shot 33 (collision time 6.11 ns, mean velocity 190 km/s, areal density 0.47 g/cm²) without explaining how the model handles a target radius it was never trained on. This is an internal inconsistency that affects the validity of the surrogate's prediction for shot 33. Please clarify whether target radius is part of the input features, and if so, correct the architecture description; if not, explain why the model can be applied to a 560 µm target.
  4. [Equation (1) and Section 5.1] The calibration loss function includes per-feature weights α_F and sums over multiple implosion features F, but the text immediately states that 'the implosion feature we employed here is the collision time tcol'. If only tcol is used in calibration, then the vector of weights α_F and the sum over features are not operational, and the calibration does not constrain velocity, areal density, or maximum density. This means the calibrated η is tuned specifically to match collision timing; the subsequent agreement of the simulated velocity and density with the (unmeasured) experimental values cannot be inferred from the calibration itself. The authors should state clearly which features enter the calibration and, if only tcol is used, discuss how this limited calibration supports the use of η for other features.
minor comments (4)
  1. [Figure 3] The boxplot residuals would be more informative if the units and the definition of the 'median residual reduced by 88.2% on average' were specified; as written, it is unclear whether the reduction is in absolute residual, squared residual, or a normalized metric.
  2. [Section 4.1] The stability constraints of the PID inverse mapping are described qualitatively ('the sampling ranges of the PID dataset should be strictly limited within the range of the original dataset'); please state the concrete procedure for enforcing this limitation, since the quality of the PID-generated samples depends on it.
  3. [Figure 5] The axis labels in Figure 5 appear garbled with unicode escape sequences in the preprint; please ensure the mathematical symbols (ρR, Vmean, tcol, ρmax) render correctly in the final version.
  4. [Section 5.1, Section 6] The sentence 'In similar with the calibrations on NIF [12]' is ungrammatical and should be rephrased; also, in Section 6, 'the transformer-based surrogate model aligns well with the simulation and experimental results' should read 'with the simulation and experimental results' or 'with the simulations and experimental results'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the effective absorption factor is calibrated on three shots, while the headline collision-time match for shot 33 is a held-out prediction; the surrogate is validated against independent MULTI-IFE simulations.

full rationale

The paper's derivation chain is not circular. The effective laser absorption rate η is explicitly a calibration parameter: Section 5.1 states that 'In the calibration regarding the collision moments, we use 3 shots from the DCI-R10 experiment to perform the calibration, including shots 22, 24, and 37,' and the resulting η≈65% is then applied to predict a different, held-out shot: 'Fig. 7 shows the prediction of a typical shot of the DCI-R10 implosion experiment... For shot 33.' Thus the claimed agreement for shot 33 (experimental collision time 6.17 ns versus simulated 6.11 ns) is a genuine transfer test rather than an in-sample fit. The inferred velocity (≈195 km/s) and density (≈117 g/cc) are outputs of the calibrated simulation, not fit targets; they could in principle disagree with future measurements, and the paper itself qualifies them by noting that 'the collision of the plasma jets in the DCI scheme is intrinsically two-dimensional' and that accurate density/temperature distributions require 2D simulations. The surrogate MULTI-Net is trained on MULTI-IFE outputs and is assessed against separate simulation test/validation sets (Fig. 3, R²≈0.999; Fig. 5, PID error reduction), so using it as a fast replacement for the hydrodynamic code is not circular. The constant-η assumption is an untested modeling assumption that limits external validity, but that is a correctness risk, not a circularity. Self-citations to the DCI scheme and streak-camera diagnostics are domain background and are not used to force the prediction. No equation or fitted parameter is reduced by construction to the claimed result.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central predictions rest on a 1D hydro code, a constant fitted absorption factor, a three-shot calibration, and a single held-out comparison; the inferred density and velocity are not independently measured.

free parameters (2)
  • effective laser absorption rate eta = about 65% (64.5% for MULTI-IFE, 65.5% for MULTI-Net)
    Fitted by minimizing loss between simulated and measured collision times over DCI-R10 shots 22, 24, and 37; then used for all subsequent predictions.
  • per-feature loss weights alpha_F in calibration = not specified in manuscript
    Defined in the calibration loss in Section 5.1 but no numerical values are given; the text uses only collision time in the calibration, so these weights are effectively chosen by hand.
assumptions (5)
  • domain assumption The 1D MULTI-IFE code captures the essential implosion dynamics of DCI-R10 when driven by an effective laser waveform.
    Used throughout Section 5; the paper acknowledges real collisions are intrinsically 2D in Section 5.2, so this assumption is load-bearing.
  • domain assumption The effective laser absorption rate eta is a constant, independent of laser power and shot-to-shot variations.
    Stated in Section 5.1: 'the energy absorption rate eta_theta is assumed to be a constant independent of the laser power.'
  • ad hoc to paper The 3-shot calibration set (shots 22, 24, 37) is representative, and a single fitted eta transfers to shot 33.
    No statistical justification beyond the observed agreement; prediction on shot 33 is one held-out comparison.
  • domain assumption The x-ray streak camera measurement of collision time and confinement duration is reliable and correctly corresponds to simulation collision time.
    Used as ground truth for calibration and validation; camera has 130 ps temporal resolution, comparable to the claimed 0.1 ns accuracy.
  • domain assumption The PID-generated laser waveforms stay within the valid range of the original dataset, keeping the inverse mapping stable.
    Explicitly noted in Section 4.1 as a stability constraint.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Predictive Hydrodynamic Simulations for Laser Direct-drive Implosion Experiments via Artificial Intelligence." pith.science (2026). https://pith.science/paper/U6PO72IP

@misc{pith2026250716227,
  author       = {Pith},
  title        = {Pith review of: Predictive Hydrodynamic Simulations for Laser Direct-drive Implosion Experiments via Artificial Intelligence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U6PO72IP}},
  note         = {Machine review of arXiv:2507.16227}
}
read the original abstract

This work presents predictive hydrodynamic simulations empowered by artificial intelligence (AI) for laser driven implosion experiments, taking the double-cone ignition (DCI) scheme as an example. A Transformer-based deep learning model MULTI-Net is established to predict implosion features according to laser waveforms and target radius. A Physics-Informed Decoder (PID) is proposed for high-dimensional sampling, significantly reducing the prediction errors compared to Latin hypercube sampling. Applied to DCI experiments conducted on the SG-II Upgrade facility, the MULTI-Net model is able to predict the implosion dynamics measured by the x-ray streak camera. It is found that an effective laser absorption factor about 65\% is suitable for the one-dimensional simulations of the DCI-R10 experiments. For shot 33, the mean implosion velocity and collided plasma density reached 195 km/s and 117 g/cc, respectively. This study demonstrates a data-driven AI framework that enhances the prediction ability of simulations for complicated laser fusion experiments.

Figures

Figures reproduced from arXiv: 2507.16227 by the authors.

Figure 1
Figure 1. (b) Sampling Methods Latin hypercube sampling Physics-Informed Decoder (d) Experiment Laser SG-II Upgrade Facility (e) Experiment Results MULTI-IFE (c) Deep Learning Model: MULTI-Net Input： Laser Waveform and Target Input Features Simulation Outputs 𝒗𝐢𝐦𝐩 𝝆𝑹 𝑰𝑭𝑨𝑹 𝝆𝐦𝐚𝐱 𝒕𝒄𝒐𝒍 (a) Hydrodynamic Simulations Calibration and Prediction Implosion Features Experiment Target Output Layers Layer Norm Multi￾Head Attention Feed Fo… view at source ↗
Figure 2
Figure 2. (1) Input Embedding: mapping sequence data into [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 2
Figure 2. Positional Encoding 𝑣imp 𝜌𝑅 𝐼𝐹𝐴𝑅 𝜌max ⋯ Linear Embed ding Input Features Implosion Features Input Embedding Layer Norm Multi￾Head Attention Feed Forward Layer Residual Connection Transformer Encoder Layer × 𝑁 Attention Pooling Prediction Head Output Layers Residual Connection Residual Connection (1) (2) (3) (4) (5) (6) (7) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: Comparison of the regress ability of the MULTI-Net model with MLP and Transformer architecture for implosion velocity and collision time. The simulated and regressed values are plot with reference line y = x . The boxplots display the interquartile range with the media…
Figure 4
Figure 4. Figure 4: 𝑃1 𝑃2 𝑃3 𝑃4 𝑃5 𝐼1 𝐼2 𝐼3 1. Train decoder with original dataset: 2.Generate implosion features by grid sampling method 3.Predict laser power by decoder 4.Calculate real implosion feature with MULTI-IFE Output Input [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: shows the prediction ability of the MULTI-Net surrogate model on the validation set. The MULTI-Net model is trained on datasets of the same size with the same number of model parameters. It is shown that the models using the PID datasets show better generalization abil…
Figure 6
Figure 6. Figure 6: The effective laser absorption rate calibrated by DCI-R10 experimental results. (MAE: Mean Absolute Error) 5.2. Prediction of DCI Experimental Implosion Dynamics For efficient prediction of experimental results, the optimal MULTI-Net model is combined with the calibrat…
Figure 7
Figure 7. Figure 7: (b) (c) (a) [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Sign in 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. Co4ICF: Co-evolving Physics-Informed Surrogate and RL-based Pulse Optimizer for Inertial Confinement Fusion

    cs.AI 2026-07 conditional novelty 6.0 of 10

    A co-evolving physics-informed surrogate and PPO optimizer reaches 146% (1D) and 247% (2D) normalized ICF yield versus baseline without 2D training data.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    Laser compression of matter to super-high densities: Thermonuclear (CTR) applications

    Nuckolls, John and Wood, Lowell and Thiessen, Albert and Zimmerman, George. “Laser compression of matter to super-high densities: Thermonuclear (CTR) applications”, Nature, 239(5368), 139–142 (1972)

    work page 1972

  2. [2]

    Achievement of target gain larger than unity in an inertial fusion experiment

    Abu-Shawareb, H et al. “Achievement of target gain larger than unity in an inertial fusion experiment”, Physical review letters, 132(6), 065102 (2024)

    work page 2024

  3. [3]

    Demonstration of hot-spot fuel gain exceeding unity in direct-drive inertial confinement fusion implosions

    Williams, CA et al. “Demonstration of hot-spot fuel gain exceeding unity in direct-drive inertial confinement fusion implosions”, Nature Physics , 20(5), 758–764 (2024)

    work page 2024

  4. [4]

    Nagatomo, H and Johzaki, T and Takizawa, R and Fujioka, S. “Formation of high areal density core using an Predictive Hydrodynamic Simulations for Laser Direct-drive Implosion Experiments via AI 7 efficient and robust implosion method for fast ignition”, Nuclear Fusion, 64(10), 106019 (2024)

    work page 2024

  5. [5]

    Physics of porous materials under extreme laser-generated conditions

    Tikhonchuk, VT and Weber, S. “Physics of porous materials under extreme laser-generated conditions”, Matter and Radiation at Extremes, 9(3), (2024)

    work page 2024

  6. [6]

    Focused energy, a new approach towards inertial fusion energy

    Ditmire, T et al. “Focused energy, a new approach towards inertial fusion energy”, Journal of Fusion Energy, 42(2), 27 (2023)

    work page 2023

  7. [7]

    Double-cone ignition scheme for inertial confinement fusion

    Zhang, Jie et al. “Double-cone ignition scheme for inertial confinement fusion”, Philosophical Transactions of the Royal Society A, 378(2184), 20200015 (2020)

    work page 2020

  8. [8]

    Machine-learning guided optimization of laser pulses for direct-drive implosions

    Wu, Fuyuan et al. “Machine-learning guided optimization of laser pulses for direct-drive implosions”, High Power Laser Science and Engineering, 10, e12 (2022)

    work page 2022

Show all 22 references
  1. [9]

    Comparison of the evolution of Rayleigh–Taylor instability during the coasting phase of the central ignition and the double-cone ignition schemes

    Lei, YY and Wu, FY and Ramis, R and Zhang, J. “Comparison of the evolution of Rayleigh–Taylor instability during the coasting phase of the central ignition and the double-cone ignition schemes”, Physics of Plasmas, 31(1), (2024)

    2024

  2. [10]

    Diagnosing the fast-heating process of the double-cone ignition scheme with X-ray spectroscopy

    Dai, Yu et al. “Diagnosing the fast-heating process of the double-cone ignition scheme with X-ray spectroscopy”, High Power Laser Science and Engineering , 12, e50 (2024)

    2024

  3. [11]

    Observation of the colliding process of plasma jets in the double-cone ignition scheme using an x-ray streak camera

    Liu, Zhengdong et al. “Observation of the colliding process of plasma jets in the double-cone ignition scheme using an x-ray streak camera”,Physics of Plasmas, 31(4), (2024)

    2024

  4. [12]

    Data-driven prediction of scaling and ignition of inertial confinement fusion experiments

    Ga ffney, Jim A et al. “Data-driven prediction of scaling and ignition of inertial confinement fusion experiments”, Physics of Plasmas, 31(9), (2024)

    2024

  5. [13]

    Transfer learning to model inertial confinement fusion experiments

    Humbird, Kelli D and Peterson, Jayson Luc and Spears, BK and McClarren, Ryan G. “Transfer learning to model inertial confinement fusion experiments”, IEEE Transactions on Plasma Science, 48(1), 61–70 (2019)

    2019

  6. [14]

    Transfer learning driven design optimization for inertial confine- ment fusion

    Humbird, Kelli D and Peterson, Jayson Luc. “Transfer learning driven design optimization for inertial confine- ment fusion”, Physics of Plasmas, 29(10), (2022)

    2022

  7. [15]

    Tripled yield in direct- drive laser fusion through statistical modelling

    Gopalaswamy, Varchas et al. “Tripled yield in direct- drive laser fusion through statistical modelling”, Nature, 565(7741), 581–586 (2019)

    2019

  8. [16]

    How numerical simulations helped to achieve breakeven on the NIF

    Marinak, MM et al. “How numerical simulations helped to achieve breakeven on the NIF”, Physics of Plasmas , 31(7), (2024)

    2024

  9. [17]

    Hybrid optimization of laser-driven fusion targets and laser profiles

    Li, Z and Zhao, ZQ and Yang, XH and Zhang, GB and Ma, YY and Xu, H and Wu, FY and Shao, FQ and Zhang, J. “Hybrid optimization of laser-driven fusion targets and laser profiles”, Plasma Physics and Controlled Fusion , 66(1), 015010 (2023)

    2023

  10. [18]

    MULTI- IFE—A one-dimensional computer code for Inertial Fusion Energy (IFE) target simulations

    Ramis, Rafael and Meyer-ter-Vehn, J ¨urgen. “MULTI- IFE—A one-dimensional computer code for Inertial Fusion Energy (IFE) target simulations”, Computer Physics Communications, 203, 226–237 (2016)

    2016

  11. [19]

    Transformer-powered sur- rogates close the ICF simulation-experiment gap with extremely limited data

    Olson, Matthew L et al. “Transformer-powered sur- rogates close the ICF simulation-experiment gap with extremely limited data”, Machine Learning: Science and Technology, 5(2), 025054 (2024)

    2024

  12. [20]

    Attention is all you need

    Vaswani, Ashish et al. “Attention is all you need”, Advances in Neural Information Processing Systems, 30, (2017)

    2017

  13. [21]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Devlin, Jacob et al. “Bert: Pre-training of deep bidirectional transformers for language understanding”, In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics, 4171–4186 (2019)

    2019

  14. [22]

    A comparison of three methods for selecting values of input variables in the analysis of output from a computer code

    McKay, Michael D et al. “A comparison of three methods for selecting values of input variables in the analysis of output from a computer code”,Technometrics, 42(1), 55–61 (2000)

    2000

Pith tools

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