Co4ICF: Co-evolving Physics-Informed Surrogate and RL-based Pulse Optimizer for Inertial Confinement Fusion

Fuyuan Wu; Jiatong Zhao; Junchi Yan; Tengyue Zhang; Yuhan Wang

REVIEW 2 major objections 5 minor 25 references

Co4ICF: Co-evolving Physics-Informed Surrogate and RL-based Pulse Optimizer for Inertial Confinement Fusion

T0 review · 2 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read A co-evolving physics-informed surrogate and RL pulse optimizer raises ICF fusion yield to 146% of baseline in 1D MULTI and 247% under direct 2D MULTI evaluation with no 2D training.

desk verdict Solid engineering paper that fixes static-surrogate OOD exploitation in ICF pulse design with a dual-frequency loop, public data, and budget-matched ablations; the 2D yield number is a single post-hoc trajectory and needs seeds. read the letter →

arxiv 2607.10366 v1 pith:ME2CCFLE submitted 2026-07-11 cs.AI

Jiatong Zhao , Tengyue Zhang , Yuhan Wang , Fuyuan Wu , Junchi Yan This is my paper

classification cs.AI

keywords InertialConfinementFusionSurrogateModelingReinforcementLearningPulseOptimizationPhysics-InformedNeuralNetworksMULTI-IFEPPODistributionShift

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

Offline neural surrogates for inertial confinement fusion fail once an optimizer drives laser pulses outside the training distribution and the model starts rewarding non-physical shapes. This paper argues that the same distribution shift can be turned into a training signal. Co4ICF runs a PPO policy against a cheap physics-regularized surrogate at high frequency, then periodically re-labels policy samples with the real MULTI radiation-hydrodynamics code and fine-tunes the surrogate on those on-policy trajectories. In the 1D MULTI loop the method reaches 146.1% of the designed-pulse baseline yield; the same pulse, checked post-hoc in 2D MULTI with no 2D data ever used, reaches 246.9%. Budget-matched ablations show that simply adding the same number of offline samples does not close the gap, so the co-evolution itself appears to be doing essential work.

What carries the argument

Dual-frequency co-evolving loop: high-frequency PPO updates treat a physics-informed non-autoregressive surrogate as a fast environment; every K steps a budgeted set of on-policy pulses is re-evaluated by MULTI and used to fine-tune the surrogate, correcting extrapolation errors as the policy shifts the input distribution.

What would settle it

Re-run the budget-matched static-surrogate ablation with identical total MULTI labels drawn from the original offline distribution and the same PPO optimizer; if its final direct 2D-MULTI normalized yield matches or exceeds Co4ICF’s reported 246.9%, the co-evolving mechanism is not required for the claimed gains.

Watch

Extended reading notes

Core claim

Optimizer-induced out-of-distribution drift, usually fatal for static ICF surrogates, becomes useful when the surrogate is fine-tuned on policy-induced trajectories: the resulting dual-frequency loop produces laser pulses whose MULTI-verified fusion yield substantially exceeds the experimental baseline and static-surrogate, Bayesian, and genetic baselines under the same labeling budget, and those pulses transfer to 2D MULTI without any 2D training or fine-tuning.

Load-bearing premise

The search is locked inside a fixed VAE-learned manifold of laser pulses; if the true high-yield shapes for the target lie outside that manifold, the optimizer cannot reach them.

Editorial extensions

If this is right

Static offline ICF surrogates can be replaced by periodically recalibrated ones that remain reliable under iterative optimization.
Pulses found entirely in 1D search can retain high performance under direct 2D radiation-hydrodynamics evaluation without any 2D training data.
High-frequency policy rollouts can be accelerated by roughly three orders of magnitude (≈990×) while still producing MULTI-verified yield gains.
Budget-matched comparisons imply the gain is not explained by extra simulation data alone, so the on-policy recalibration schedule itself is a design lever.
The released large-scale MULTI-IFE dataset enables shared benchmarking of future surrogate–optimizer methods.

Reading between the lines

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

Other high-cost design loops that currently freeze a learned simulator (molecular design, materials discovery) may gain more from the same on-policy recalibration schedule than from simply collecting more offline data.
The jump from 146% (1D) to 247% (2D) normalized yield suggests some 1D-discovered structures are rewarded even more under 2D physics; a 2D-calibrated co-evolving loop would test whether the ceiling is higher still.
Because the decoder is frozen, occasional manifold expansion or decoder fine-tuning is a direct test of whether higher yields exist outside the current feasible set.

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.

Desk Editor's Note

read the letter

The useful thing here is a clean, implemented fix for the failure mode everyone already knows about: offline ICF surrogates get exploited once the optimizer leaves the training distribution. They couple a physics-regularized NAR Transformer surrogate with PPO in a fixed VAE pulse manifold, update the policy at high frequency against the surrogate, and fine-tune the surrogate at low frequency on MULTI-labeled on-policy pulses. That dual-frequency loop is the concrete contribution; the rest (Dyna-style co-evolution, hinge physics losses, PPO pulse shaping) has clear precedents they cite.

What they do well is the evidence package inside the MULTI world. Table 3 and Fig. 9 give a frozen-surrogate control, a budget-matched “enlarged static” ablation that draws the same extra labels from the offline distribution, and BO/GA baselines under identical labeling budgets. Co4ICF lands at 146.1% normalized 1D yield and 246.9% on a direct post-hoc 2D-MULTI evaluation of the same pulse, with no 2D data ever used in training or selection. Physics constraints improve OOD peak metrics (Table 1). They ship ~90k MULTI trajectories and code. That is real, reproducible engineering value for anyone running expensive radiation-hydro loops.

Soft spots, in proportion. The stress-test is right on the statistics: every headline yield is a single point with no seeds, no error bars, and no distribution over final designs. The reward is a ρRT proxy; the 2D number is one trajectory through a stochastic optimizer. If seed variance is large, the causal claim that co-evolution (not luck or extra data) produced the gap over the 174.8% enlarged baseline weakens. The fixed VAE manifold is an acknowledged structural limit, not a hidden flaw. MULTI itself cannot capture 3-D instabilities; they say so. None of these sink the 1D diagnostic or the public dataset.

This is for people who actually design pulses or build surrogates for expensive simulators. It is not a physics breakthrough and does not claim to be. I would send it to referees; they will ask for multi-seed runs and uncertainty on the final yields, which is fair. Worth engaging, citing for the dataset and the ablation design, and bringing to a methods reading group.

Referee Report

2 major / 5 minor

Summary. The paper presents Co4ICF, a dual-frequency co-evolving loop that couples a physics-informed non-autoregressive surrogate (Transformer encoder with mass/momentum/energy hinge constraints derived from a steady-ablation model) with a PPO pulse optimizer that acts in a fixed VAE latent manifold. High-frequency PPO updates query the surrogate as a cheap environment; every K steps, a budgeted set of on-policy pulses is re-labeled with 1D-MULTI and used to fine-tune the surrogate. On a fixed target, the method reports 146.1% normalized yield after 1D-MULTI re-evaluation of the optimization trajectory and 246.9% when the same final pulse is evaluated post-hoc in 2D-MULTI (no 2D data used for training or selection). Budget-matched ablations (static surrogate, enlarged static surrogate trained on the same extra sample count drawn from the offline distribution, BO, GA) and a surrogate–MULTI peak-error diagnostic (Fig. 9) are offered as evidence that the gain is not explained by extra data alone. A large MULTI-IFE dataset and code are released.

Significance. If the reported gains and the causal role of on-policy refinement hold under modest statistical scrutiny, the work supplies a concrete, reproducible recipe for simulation-informed design under limited high-fidelity budgets: treat optimizer-induced distribution shift as a training signal rather than only as a failure mode. The combination of a physics-regularized NAR surrogate, a fixed VAE manifold that enforces feasibility, dual-frequency updates, and a public ~90k-shot MULTI dataset is a useful contribution to AI-for-fusion and more generally to offline model-based optimization. The ~990× inner-loop speedup and the post-hoc 1D-to-2D transfer (even if only a single trajectory) are practically relevant. Strengths that should be credited include the budget-matched enlarged-static control, the explicit OOD diagnostic in Fig. 9, the physics-constraint ablations in Table 1, and the open dataset/code release.

major comments (2)

Table 3 and Fig. 1 report single-point normalized yields (146.1% 1D, 246.9% 2D) with no error bars, no multi-seed PPO runs, and no distribution over final designs. The 2D number is a pure post-hoc evaluation of one trajectory whose training reward is the ρRT proxy in Eq. (1), not yield. Without at least a small multi-seed comparison of W/Update vs. W/O Update (Enlarged) under the same labeling budget M, the causal attribution of the 246.9% vs. 174.8% gap to the co-evolving mechanism remains under-supported. This is load-bearing for the abstract’s strongest claim.
The optimization is confined to a fixed VAE decoder pretrained on Real+PPCA data (Stage I, §3.1; Limitations). The paper correctly notes that high-yield shapes outside this manifold are unreachable, yet provides no quantitative characterization of manifold coverage relative to the high-yield region (e.g., reconstruction error of known high-yield pulses, or a controlled comparison against unconstrained or less-constrained search). Without that, it is hard to judge how much of the reported gain is limited by the manifold versus enabled by co-evolution.

minor comments (5)

Eq. (1) uses a ρRT proxy for training while Table 3 reports MULTI-verified fusion yield; a short scatter or correlation plot of ρRT vs. yield on the optimized set would make the proxy–objective relationship transparent.
Table 1 cell coloring and the AR/NAR comparison note are helpful, but the AR model is never given physics constraints; a one-sentence justification in the caption would avoid the appearance of an unfair comparison.
Hyperparameters λ_m, λ_p, λ_E, α, K, M, d_z, and ε are free but not tabulated with values or sensitivity; a short appendix table would aid reproducibility.
Fig. 8 t-SNE of policy samples is qualitative; reporting a simple distance-to-Real-cluster statistic over training stages would strengthen the distribution-shift narrative.
Minor notation: the policy is π_ω in §3.1 and π_θ in §3.4/Appendix C; unify the subscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: reported yields are MULTI/2D-MULTI evaluations of optimized pulses, not quantities forced by the surrogate objective or by self-citation.

full rationale

Co4ICF is an empirical co-evolution pipeline (offline MULTI pretraining → high-frequency PPO on a physics-regularized NAR surrogate → low-frequency MULTI relabeling of on-policy pulses). The headline numbers 146.1% (1D) and 246.9% (post-hoc 2D) are defined as MULTI-verified fusion yield divided by an external experimental baseline pulse’s MULTI yield (Eq. 1 reward is only a ρRT proxy used during search; Table 3 and Fig. 1 report simulator yields). Budget-matched ablations (W/O Update Static vs Enlarged vs W/ Update) retrain on the same extra sample count drawn from the offline distribution rather than the policy, so the co-evolution attribution is comparative, not definitional. Physics-informed losses are one-sided hinge penalties from a steady-ablation upper-bound model (Appendix B, external Atzeni formula); they regularize trajectory consistency and do not fit the reported yield percentages. Related-work self-citations ([22]–[24]) are background on ICF surrogates/optimization and are not load-bearing uniqueness theorems. The fixed VAE manifold is an explicit search-space limitation, not a circular derivation. No step reduces a claimed first-principles prediction to its own fitted inputs by construction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central performance claim rests on MULTI as ground truth, on the VAE manifold being sufficiently expressive, on the steady-ablation inequalities being useful regularizers, and on a handful of loss and PPO hyper-parameters. No new physical entities are postulated; the framework itself is the main invented construct.

free parameters (4)

physics-loss weights λ_m, λ_p, λ_E
Trade-off coefficients between data loss and the three hinge penalties; chosen during surrogate pretraining and not derived from first principles.
energy-penalty coefficient α
Scales the soft energy-budget term in the PPO reward (Eq. 1); set by the authors.
surrogate fine-tune interval K and labeling budget M
Control how often and how many on-policy MULTI labels are collected; directly affect the co-evolution dynamics.
VAE latent dimension d_z and PPO clip ε
Architectural and optimization hyper-parameters that define the searchable action space and update size.

assumptions (4)

domain assumption MULTI-IFE (1D and 2D) trajectories are sufficiently faithful ground truth for the reported yield comparisons.
All rewards, fine-tuning labels and final metrics are MULTI-derived; Limitations section notes that MULTI cannot capture 3-D Rayleigh–Taylor asymmetries.
domain assumption Steady-ablation mass, momentum and energy upper bounds (Appendix B) are valid one-sided constraints that improve OOD behavior.
Derived from simplified rocket-effect formulas; used as soft hinge penalties rather than hard equalities.
domain assumption Peak ρRT is a useful proxy reward that correlates with fusion yield on the optimized pulses.
Stated in Section 3.4; final tables report MULTI yield, not the proxy.
ad hoc to paper The fixed VAE decoder maps the latent prior onto a manifold that contains high-performing feasible pulses.
Search never leaves the decoder image; Limitations explicitly flags the risk of missing shapes outside the manifold.

invented entities (2)

Co4ICF dual-frequency co-evolving loop independent evidence
purpose: Jointly update PPO actor at high frequency against the surrogate and fine-tune the surrogate at low frequency on MULTI-labeled on-policy samples.
The concrete scheduling and coupling of the two modules is the paper’s central methodological construct; independent evidence is the ablation gap versus static and enlarged-static surrogates.
Differentiable mass-from-index operator plus squared-hinge mass/momentum/energy losses
purpose: Inject approximate conservation inequalities into the non-autoregressive surrogate without requiring equality constraints.
Specific formulation in Appendix B; improves OOD metrics in Table 1 but is an engineering choice, not a new physical law.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Co4ICF: Co-evolving Physics-Informed Surrogate and RL-based Pulse Optimizer for Inertial Confinement Fusion." pith.science (2026). https://pith.science/paper/ME2CCFLE

@misc{pith2026260710366,
  author       = {Pith},
  title        = {Pith review of: Co4ICF: Co-evolving Physics-Informed Surrogate and RL-based Pulse Optimizer for Inertial Confinement Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ME2CCFLE}},
  note         = {Machine review of arXiv:2607.10366}
}

read the original abstract

Offline-trained surrogates for Inertial Confinement Fusion (ICF) suffer a well-known failure mode that iterative optimizers drive inputs into out-of-distribution (OOD) regions where predictions become unreliable. Here we present Co4ICF, a co-evolving framework that couples a physics-informed surrogate with a PPO-based pulse optimizer. The surrogate is iteratively fine-tuned on policy-induced trajectories, correcting extrapolation errors as the optimizer shifts the input distribution; the optimizer queries this evolving surrogate as a fast environment. In the 1D MULTI environment, Co4ICF achieves 146.1% normalized yield based on current laser design baseline; as a post-hoc cross-fidelity check, the optimized pulse further attains 246.9% normalized yield when directly evaluated in 2D-MULTI without any 2D training or fine-tuning. Budget-matched ablations support that the gains are not explained solely by additional simulation data and are consistent with the co-evolving mechanism playing a key role. We release a large-scale MULTI-IFE simulation dataset to support future benchmarking.

Figures

Figures reproduced from arXiv: 2607.10366 by the authors.

**Figure 2.** Co4ICF framework. Stage I (Pretrain): A VAE decoder learns a latent representation of laser pulses; the surrogate is [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Inputs and outputs of the MULTI-IFE simulation. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 5.** Figure 5: Dataset composition. Real: 701 measured pulses; Clean: 10k VAE-augmented in-distribution samples; Other: 80k OOD pulses from Fourier noise and PPCA sampling. 3.3 Surrogate with Physics-Informed Regularization The surrogate maps laser pulse inputs to time-resolved implo…

**Figure 6.** Figure 6: Physics-Informed Loss. The surrogate predicts tra [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

**Figure 7.** Figure 7: 𝐿2 loss ratio during surrogate pretraining. Transformer encoder is the strongest NAR surrogate. It offers the strongest balance between accuracy and robustness among NAR models. With physics constraints, it has the lowest OOD peak-𝑥 MAE (6.302), the lowest OOD peak-𝑦 M…

**Figure 8.** Figure 8: t-SNE visualization of laser pulses during Co4ICF [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

**Figure 9.** Figure 9: Surrogate–MULTI peak error over PPO training [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

**Figure 10.** Figure 10: t-SNE visualization of input laser pulses across [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 1 canonical work pages

[1]

Rushil Anirudh, Jayaraman J Thiagarajan, Peer-Timo Bremer, and Brian K Spears. 2020. Improved surrogates in inertial confinement fusion with mani- fold and cycle consistencies.Proceedings of the National Academy of Sciences 117, 18 (2020), 9741–9746

2020
[2]

2004.The Physics of Inertial Fusion: Beam–Plasma Interaction, Hydrodynamics, Hot Dense Matter

Stefano Atzeni and Jürgen Meyer ter Vehn. 2004.The Physics of Inertial Fusion: Beam–Plasma Interaction, Hydrodynamics, Hot Dense Matter . Oxford University Press, Oxford, UK. doi:10.1093/acprof:oso/9780198562641.001.0001

doi:10.1093/acprof:oso/9780198562641.001.0001 2004
[3]

Francesco Capuano, Davorin Peceli, and Gabriele Tiboni. 2025. Shaping Laser Pulses with Reinforcement Learning.arXiv preprint arXiv:2503.00499 (2025)

arXiv 2025
[4]

Jonas Degrave, Federico Felici, Jonas Buchli, Michael Neunert, Brendan Tracey, Francesco Carpanese, Timo Ewalds, Roland Hafner, Abbas Abdolmaleki, Diego de Las Casas, et al. 2022. Magnetic control of tokamak plasmas through deep reinforcement learning.Nature 602, 7897 (2022), 414–419

2022
[5]

Rahman Ejaz, Varchas Gopalaswamy, A Lees, C Kanan, D Cao, and R Betti. 2024. Deep learning-based predictive models for laser direct drive at the Omega Laser Facility.Physics of Plasmas 31, 5 (2024)

2024
[6]

V Gopalaswamy, A Lees, R Ejaz, CA Thomas, TJB Collins, KS Anderson, W Ebmeyer, and R Betti. 2025. Automated and highly parallelized Bayesian op- timization scheme for direct drive fusion experiments on OMEGA.Physical Review Research 7, 1 (2025), 013009

2025
[7]

Kelli Denise Humbird. 2019. Machine learning guided discovery and design for inertial confinement fusion. Ph. D. Dissertation. Texas A&M University

2019
[8]

Kelli D Humbird, J Luc Peterson, J Salmonson, and Brian K Spears. 2021. Cog- nitive simulation models for inertial confinement fusion: Combining simulation and experimental data.Physics of Plasmas 28, 4 (2021)

2021

Show all 25 references

[9]

Omar A Hurricane. 2024. How ignition and target gain> 1 were achieved in inertial fusion.High Energy Density Physics 53 (2024), 101157

2024
[10]

George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. 2021. Physics-informed machine learning.Nature Reviews Physics 3, 6 (2021), 422–440

2021
[11]

Minsu Kim, Jiayao Gu, Ye Yuan, Taeyoung Yun, Zixuan Liu, Yoshua Bengio, and Can Chen. 2025. Offline Model-Based Optimization: Comprehensive Review. arXiv:2503.17286 [cs.LG] https://arxiv.org/abs/2503.17286

2025
[12]

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

2023
[13]

Shulei Liu, Handing Wang, Wei Peng, and Wen Yao. 2024. Surrogate-assisted evo- lutionary algorithms for expensive combinatorial optimization: a survey.Com- plex & Intelligent Systems 10, 4 (Aug. 2024), 5933–5949. doi:10.1007/s40747-024- 01465-5

2024 doi
[14]

Matthew L Olson, Shusen Liu, Jayaraman J Thiagarajan, Bogdan Kustowski, Weng-Keen Wong, and Rushil Anirudh. 2024. Transformer-powered surrogates close the ICF simulation-experiment gap with extremely limited data.Machine Learning: Science and Technology 5, 2 (2024), 025054

2024
[15]

Maziar Raissi, Paris Perdikaris, and George E Karniadakis. 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 (2019), 686–707

2019
[16]

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

2016
[17]

Brian K Spears, Scott Brandon, Dan T Casey, John E Field, Jim A Gaffney, Kelli D Humbird, Andrea L Kritcher, Michael KG Kruse, Eugene Kur, Bogdan Kustowski, et al. 2025. Predicting fusion ignition at the National Ignition Facility with physics-informed deep learning.Science 38...

2025
[18]

Richard S Sutton. 1991. Dyna, an integrated architecture for learning, planning, and reacting.ACM SIGART Bulletin 2, 4 (1991), 160–163

1991
[19]

Tao Tao, Guannan Zheng, Qing Jia, Rui Yan, and Jian Zheng. 2023. Laser pulse shape designer for direct-drive inertial confinement fusion implosions.High Power Laser Science and Engineering 11 (2023), e41

2023
[20]

Brandon Trabucco, Xinyang Geng, Aviral Kumar, and Sergey Levine. 2022. Design-Bench: Benchmarks for Data-Driven Offline Model-Based Optimization. arXiv:2202.08450 [cs.LG] https://arxiv.org/abs/2202.08450

2022 arXiv
[21]

Jingyi Wang, N Chiang, Andrew Gillette, and J Luc Peterson. 2024. A multi- fidelity Bayesian optimization method for inertial confinement fusion design. Physics of Plasmas 31, 3 (2024)

2024
[22]

Zixu Wang, Yuhan Wang, Junfei Ma, Fuyuan Wu, Junchi Yan, Xiaohui Yuan, Zhe Zhang, and Jie Zhang. 2025. Predictive Hydrodynamic Simulations for Laser Direct-drive Implosion Experiments via Artificial Intelligence.arXiv preprint arXiv:2507.16227 (2025)

2025 arXiv
[23]

S Wei, F Wu, Y Zhu, J Yang, L Zeng, X Li, and J Zhang. 2024. A Machine Learning Method for the Optimization Design of Laser Pulse in Fast Ignition Simulations. Journal of Fusion Energy 43, 1 (2024), 6

2024
[24]

Fuyuan Wu, Xiaohu Yang, Yanyun Ma, Qi Zhang, Zhe Zhang, Xiaohui Yuan, Hao Liu, Zhengdong Liu, Jiayong Zhong, Jian Zheng, et al. 2022. Machine-learning guided optimization of laser pulses for direct-drive implosions.High Power Laser Science and Engineering 10 (2022), e12

2022
[25]

Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2022. Are Transformers Effective for Time Series Forecasting? arXiv:2205.13504 [cs.AI] https://arxiv. org/abs/2205.13504 A Dataset and Data Augmentation This section describes the dataset composition and the three data augmentat...

2022 arXiv

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.

[1] [1]

Rushil Anirudh, Jayaraman J Thiagarajan, Peer-Timo Bremer, and Brian K Spears. 2020. Improved surrogates in inertial confinement fusion with mani- fold and cycle consistencies.Proceedings of the National Academy of Sciences 117, 18 (2020), 9741–9746

2020

[2] [2]

2004.The Physics of Inertial Fusion: Beam–Plasma Interaction, Hydrodynamics, Hot Dense Matter

Stefano Atzeni and Jürgen Meyer ter Vehn. 2004.The Physics of Inertial Fusion: Beam–Plasma Interaction, Hydrodynamics, Hot Dense Matter . Oxford University Press, Oxford, UK. doi:10.1093/acprof:oso/9780198562641.001.0001

doi:10.1093/acprof:oso/9780198562641.001.0001 2004

[3] [3]

Francesco Capuano, Davorin Peceli, and Gabriele Tiboni. 2025. Shaping Laser Pulses with Reinforcement Learning.arXiv preprint arXiv:2503.00499 (2025)

arXiv 2025

[4] [4]

Jonas Degrave, Federico Felici, Jonas Buchli, Michael Neunert, Brendan Tracey, Francesco Carpanese, Timo Ewalds, Roland Hafner, Abbas Abdolmaleki, Diego de Las Casas, et al. 2022. Magnetic control of tokamak plasmas through deep reinforcement learning.Nature 602, 7897 (2022), 414–419

2022

[5] [5]

Rahman Ejaz, Varchas Gopalaswamy, A Lees, C Kanan, D Cao, and R Betti. 2024. Deep learning-based predictive models for laser direct drive at the Omega Laser Facility.Physics of Plasmas 31, 5 (2024)

2024

[6] [6]

V Gopalaswamy, A Lees, R Ejaz, CA Thomas, TJB Collins, KS Anderson, W Ebmeyer, and R Betti. 2025. Automated and highly parallelized Bayesian op- timization scheme for direct drive fusion experiments on OMEGA.Physical Review Research 7, 1 (2025), 013009

2025

[7] [7]

Kelli Denise Humbird. 2019. Machine learning guided discovery and design for inertial confinement fusion. Ph. D. Dissertation. Texas A&M University

2019

[8] [8]

Kelli D Humbird, J Luc Peterson, J Salmonson, and Brian K Spears. 2021. Cog- nitive simulation models for inertial confinement fusion: Combining simulation and experimental data.Physics of Plasmas 28, 4 (2021)

2021