{"id":"6d61346d-255f-48b0-bb63-ca3cb819cd41","arxiv_id":"2508.20878","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A multi-fidelity active-learning optimizer found a 25 kJ laser-direct-drive target whose hydrodynamically scaled 2 MJ simulation burns with about 217x alpha amplification, far more than the 1D-optimized design's 17x.","lead":"This paper builds a machine-learning design system that uses cheap 1D fusion simulations to guide expensive 2D simulations and find laser targets that stay stable. The best target, scaled to a 2 MJ laser, simulates as a high-gain burning implosion, while the purely 1D-optimized target fails.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"2 MJ high-gain claim rests on unverified hydro-equivalence scaling from a 25 kJ restart, not a self-consistent 2 MJ simulation.","rationale":"The reader's weakest assumption correctly identifies the hydro-scaling step as the most fragile link in the central claim. The O2D design's 2 MJ performance is not the result of a self-consistent 2 MJ simulation but of scaling a 25 kJ restart, with the authors explicitly acknowledging that hydro-equivalence would require re-optimisation they did not perform. This concern is load-bearing because the abstract and conclusions present the 2 MJ yield as confirmation of high gain. The paper is otherwise transparent and provides useful evidence: a large 1D database, calibrated NN ensembles, transfer learning, and open-source orchestration. A single full 2 MJ simulation with a re-tuned hydro-equivalent pulse would settle whether the scaled extrapolation is representative. Since the reader already issued a CONDITIONAL verdict and the concern supports that conditional status, no verdict change is needed. A secondary but related concern about single-seed statistics is also addressed by the proposed repeated-seed test.","tokens_in":793,"tokens_out":719,"duration_ms":63271,"concrete_test":"Run a full 2D 2 MJ simulation of O2D using a hydro-equivalent laser pulse and target re-optimised according to the recipe in Nora et al. (Ref. 19), initialized with the same class of ablator density perturbations, and repeat with at least 3-5 independent perturbation seeds. Compare the burn-on yield and yield amplification against Table II. If the full-implosion yield drops below ~1e19 or amplification falls far below 217, the current high-gain confirmation is an artifact of the unverified scaling assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central high-gain result (Table II: O2D gives 2.24e19 neutrons, yield amplification 217, ~15% burn fraction at 2 MJ) is obtained by hydrodynamically scaling the 25 kJ 2D simulation fields from a restart time approaching stagnation, not by simulating a 2 MJ implosion from the beginning. The authors state this explicitly in Section IV B: 'we will not perform this re-optimisation but instead apply hydro-scaling to the 2D simulations at a restart time approaching stagnation, making the assumption that a hydro-equivalent implosion can be achieved up to this time.' This is the load-bearing step. Hydro-equivalence is not guaranteed for an unmodified design: Nora et al. (Ref. 19) show that restoring hydro-equivalence requires target and pulse re-tuning (mainly exchanging ablator mass for DT ice). Without that re-optimisation, the scaled stagnation state is an extrapolation, and the 2 MJ burn-on yield/amplification is not a direct confirmation of high gain. The robustness claim is further weakened by the reported 2 MJ burn-on runs being single realizations of the randomized ablator perturbations; with ignition margin near chi_no_alpha approximately 1.72, seed-to-seed variation could plausibly cross the ignition threshold and change the yield by orders of magnitude.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an automated, multi-fidelity simulation-based design framework for laser direct drive inertial confinement fusion. The authors use ~12.5k 1D Chimera simulations to train an ensemble of MLP surrogates, transfer learned features to a much smaller 2D dataset, and then run active learning (probabilistic threshold sampling) and Bayesian optimisation (log expected improvement) over an 8-parameter, OMEGA-relevant design space at 25 kJ. The objective Y (Eq. 3a) combines a hydro-scaled, no-alpha ignition metric with an areal-density term for burn propagation. The 2D-optimised design O2D is reported to be more hydrodynamically stable than the 1D-optimised design O1D at 25 kJ. The paper then hydro-scales the 25 kJ 2D restart fields to 2 MJ and runs 2D burn-off/burn-on simulations, reporting for O2D a yield of 2.24e19 neutrons, yield amplification ~217, and ~15% burn fraction, versus 1.02e18 neutrons and amplification ~17 for O1D.","tokens_in":21705,"tokens_out":9099,"duration_ms":92236,"significance":"If the 25 kJ results are taken at face value, the paper gives a useful demonstration that a multi-fidelity surrogate, active learning, and Bayesian optimisation loop can identify a design that is more robust in 2D than a purely 1D-optimised design while using only a modest number of expensive 2D runs. The open-source orchestration package, the calibrated NN ensembles, and the explicit statement of the hydro-scaling caveat are strengths. However, the headline 2 MJ high-gain claim is not an independent confirmation: it rests on the assumption that a hydro-equivalent implosion can be achieved up to a restart time near stagnation, which the authors explicitly do not test. This makes the central claim conditional, and the conclusion as currently worded overstates the evidence. The paper is a reasonable candidate for publication after the high-gain claim is re-framed and the uncertainty of the 2 MJ results is addressed.","major_comments":[{"comment":"The 2 MJ high-gain result (O2D: 2.24e19 neutrons, yield amplification 217, ~15% burn fraction) is not a direct confirmation of high gain. The authors state: \"we will not perform this re-optimisation but instead apply hydro-scaling to the 2D simulations at a restart time approaching stagnation, making the assumption that a hydro-equivalent implosion can be achieved up to this time.\" The table footnote repeats that full hydro-equivalent re-tuning was not performed. Nora et al. (Ref. 19), cited by the authors, show that restoring hydro-equivalence requires additional target and pulse changes, mainly exchanging ablator mass for DT ice. Since the conclusion in Section VI says the framework \"correctly identified a high-gain scaled-up design,\" that claim is stronger than the simulation evidence. The stress-test concern therefore lands. I recommend either running a full 2 MJ simulation from the","section":"Section IV B / Table II"},{"comment":"The 2 MJ burn-on and burn-off values are single 2D simulations for each design, with no error bars or seed-to-seed statistics. This matters because the 25 kJ 2D value of chi_S,no_alpha for O2D is 1.72, close to the nominal ignition threshold, and the 2D simulations intentionally use randomized ablator density perturbations to represent shot-to-shot variability. A different random seed at the 2 MJ scale could plausibly change whether burn propagates and could move the yield by orders of magnitude. The robustness claim therefore needs at least a small ensemble (e.g., 3-5 perturbation realizations) at 2 MJ, or a quantitative propagation of the 25 kJ seed distribution through the scaling procedure. Without this, the 2 MJ performance numbers should be treated as a single realization, not a robust prediction.","section":"Section IV B / Table II"},{"comment":"Quantitative validation of the 2D surrogate on held-out 2D data is not reported. Fig. 1(b)(iii) shows a before/after comparison for the transfer-learned 2D model but gives no R^2, RMSE, or coverage numbers for a 2D test set. Fig. 4 compares 1D and 2D surrogates on inputs from the 1D database, not on held-out 2D simulations. Since the active learning and Bayesian optimisation decisions are driven by the 2D ensemble's mean and calibrated uncertainty, the paper should include a table or plot with held-out 2D prediction error and calibration statistics, ideally in the region around the reported optima. This would let the reader assess how much the 25 kJ and consequently 2 MJ conclusions depend on surrogate accuracy.","section":"Section II C 2 / Fig. 1"}],"minor_comments":[{"comment":"The notation d^2 Y_DT / (dt dV) is non-standard and dimensionally confusing. Please write the double integral explicitly (over time and volume) or use a clearer mixed-derivative notation.","section":"Section II A, Eq. (4c)"},{"comment":"\"2.5um\" should be \"2.5 μm\".","section":"Section II B"},{"comment":"The sentence \"First, both designs have sufficient ignition margin to ignite in 2D at the 2 MJ energy scale\" is difficult to reconcile with the O1D result, which is later described as failing to propagate burn into the dense fuel. Recommend distinguishing hot-spot ignition from propagating burn, e.g., \"sufficient margin for hot-spot ignition but insufficient confinement for propagating burn.\"","section":"Section IV B"},{"comment":"The caption says \"2D dataset\" but does not state whether the points are a held-out 2D test set or the training set. Please clarify and include quantitative performance metrics on the figure.","section":"Fig. 1(b)(iii)"},{"comment":"The phrase \"confirm the achievement of high gain\" and similar wording in the conclusion is stronger than the conditional statement in Section IV B. If the major comment about hydro-equivalence is addressed by re-framing, the abstract and conclusions should be aligned with the conditional framing.","section":"Abstract / Conclusions"}],"recommendation":"major_revision","confidential_remarks":"The paper is transparent about its main assumption, and the issues are fixable; I do not see grounds for rejection. The main risk is that the 2 MJ high-gain numbers will be quoted without the hydro-equivalence caveat, so the final version should make the conditional status prominent in the abstract and conclusions. The novelty is incremental relative to existing transfer-learning/active-learning/BO work, but the integration and demonstration on a concrete ICF problem are useful. No citation concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you care about ML-driven design for ICF. The real result is the 25 kJ demonstration: a multi-fidelity loop that takes roughly 12.5k cheap 1D simulations, transfers knowledge to 2D with about 128 expensive runs, and identifies a design (O2D) that is markedly more stable than the 1D-optimized design under seeded beam-mode and ablator-roughness perturbations. That part is internally consistent, well described, and the orchestration code (mille-feuille) is open source. Credit where due: they calibrate the NN ensemble uncertainties, compare against a GP baseline, and are explicit about what they did and did not do.\n\nThe headline 2 MJ result—O2D with yield amplification 217 and ~15% burn fraction—should not be quoted without a big caveat. It is not a self-consistent 2 MJ simulation. The authors hydro-scale the 25 kJ 2D fields from a restart near stagnation and explicitly say they do not re-optimize the pulse or target for hydro-equivalence. Nora et al. (Ref. 19) show that hydro-equivalence requires re-tuning, mainly exchanging ablator mass for DT ice. So the 2 MJ confirmation is an extrapolation, not a direct verification. The authors are honest about this in Section IV B, but the abstract and conclusions carry that number without the same emphasis, which will mislead someone skimming.\n\nThe other soft spot is statistical: the 2 MJ burn-on runs are single realizations, and chi_no_alpha ~1.72 sits close enough to the ignition threshold that seed-to-seed variation could plausibly flip the outcome. Also, the objective is built from the hydro-scaled ignition metric, so the 2 MJ outcome is partly by construction; not circular, since the burn-on runs are real simulations, but the verification is not independent of the optimization target. Repeating the 2 MJ runs with a few seeds, or better, re-optimizing at 2 MJ, would firm it up. These are moderate concerns, not fatal ones; the 25 kJ comparison is not undermined.\n\nWorth a serious referee. The method is a genuine integration of existing components, and the 25 kJ result is a plausible demonstration that 1D-to-2D transfer learning can cut the 2D simulation budget substantially. The 2 MJ claim needs re-optimization or at least seed variation before it becomes a cited high-gain prediction. I'd send it to review and ask for those additions, not desk reject it.","headline":"Worth reading for the 25 kJ multi-fidelity design loop; don't quote the 2 MJ high-gain numbers without the hydro-scaling caveat.","tokens_in":22126,"tokens_out":3335,"would_cite":true,"duration_ms":34176,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A machine-learning design loop that leans on cheap 1D simulations finds a laser-fusion target which, hydrodynamically scaled from 25 kJ to 2 MJ, burns with 217-fold yield amplification instead of collapsing under seeded instabilities.","keywords":["inertial confinement fusion","laser direct drive","multi-fidelity surrogate models","active learning","Bayesian optimisation","hydrodynamic instabilities","neural network ensemble","hydrodynamic scaling"],"falsifier":"Run the O2D design at 2 MJ with a properly hydro-equivalent re-tuned pulse and target, for example exchanging ablator mass for DT ice as hydro-equivalent ignition theory prescribes, and compare the burn-on yield against the reported 217x amplification; also fire the O2D target on a 25 kJ laser with seeded surface perturbations and compare yield and areal density with the 2D predictions. Agreement would confirm the scaled result, while a collapse toward the 1D-optimised design's 17x amplification would falsify it.","tokens_in":2053,"feed_emoji":"⚛️","tokens_out":4686,"duration_ms":103827,"temperature":0.7,"pith_summary":"The paper's aim is to make laser-direct-drive fusion design resilient to hydrodynamic instabilities without paying the full cost of multidimensional simulations. It builds a neural-network surrogate that first learns the good-design region from roughly 12,500 cheap 1D simulations, then transfers that knowledge to a small set of about 128 expensive 2D simulations that seed beam-mode and ablator-surface perturbations. Bayesian optimisation on the 2D-trained surrogate identifies an eight-parameter design at 25 kJ that is only mildly degraded by instabilities, whereas the design optimised in 1D loses its shell. Hydrodynamically scaled 2D simulations at 2 MJ, with alpha heating on, give about 2.2e19 neutrons, 217x yield amplification, and roughly 15% burn fraction for the 2D-optimised design, versus 1.0e18 neutrons, 17x amplification, and 0.9% burn for the 1D-optimised design. The core claim is that instabilities can be designed against at small scale using an information-transfer surrogate, and that the resulting target ignites and propagates burn when scaled up.","feed_headline":"Stability-first fusion target yields 217x burn amplification","feed_subtitle":"Machine learning on cheap 1D simulations finds a 25 kJ target that still burns at 2 MJ.","key_machinery":"The load-bearing mechanism is a multi-fidelity neural-network ensemble trained with transfer learning: a 25-member multilayer-perceptron ensemble learns the 1D design landscape, then all but the last two layers are frozen and the remaining layers are retrained on 2D data, so knowledge of where good designs live transfers across fidelities. Around it sits a pipeline of quasi-random initial sampling, probabilistic-threshold active learning to concentrate 2D runs in promising regions, and Bayesian optimisation with a log Expected Improvement acquisition function using the multi-fidelity surrogate at fixed fidelity. The objective function couples the chi_no_alpha ignition criterion to areal dens","core_discovery":"The central discovery is that the space of designs resilient to hydrodynamic instabilities can be learned almost entirely from 1D simulations, with a modest number of 2D simulations used as transfer data. The authors define a single scalar objective that blends the no-alpha ignition metric chi_no_alpha with post-ignition burn propagation estimated through areal density, scaled by the hydrodynamic scale factor S from 25 kJ to 2 MJ. An ensemble of 25 multilayer-perceptron surrogates trained on the 1D dataset, with all but the last two layers frozen and retrained on the 2D dataset, provides calibrated uncertainties for active learning and Bayesian optimisation. The resulting 2D-optimised design","pith_inferences":["If the hydro-equivalence assumption fails at 2 MJ, the reported O2D yield is an extrapolation; the design should be re-tuned with a full hydro-equivalent pulse and target, then re-simulated, before high gain is treated as confirmed.","The same transfer-learning loop could be pointed at 3D simulations, experimental data, or additional perturbation sources such as ice roughness and stalk shadows; each would likely shift the resilient design region the surrogate finds.","A direct 25 kJ laser experiment firing the O2D target with seeded surface perturbations would test whether the predicted resilience appears in measured yield and areal density; agreement would strengthen the scaled 2 MJ claim.","Amplifying the seeded ablator perturbation level should push the optimised design toward smaller capsules, thicker ice, and higher picket power, as the authors note; this trend is a testable prediction of the surrogate's learned physics."],"forward_implications":["A 2D-optimised design at 25 kJ is substantially more stable than the 1D-optimised one: peak inflight aspect ratio 30 versus 37, higher shell adiabat, and an intact shell at bang time.","Hydro-scaling to 2 MJ in 2D with burn-on gives the 2D-optimised design about 2.24e19 neutrons, yield amplification 217, and about 15% burn fraction, versus 1.02e18 neutrons, 17x, and 0.9% for the 1D-optimised design.","The framework locates resilient designs with only about 128 2D simulations because 1D transfer narrows the search space dramatically.","The optimiser automatically found shock-timing and stability tradeoffs, such as a higher picket power and thicker ice, without being explicitly asked to tune those quantities.","The chosen objective rewards designs that would ignite and propagate burn at a larger scale rather than merely maximising 25 kJ yield."],"supporting_citations":[{"why":"Shows multidimensional simulations are needed to capture Rayleigh-Taylor growth that degrades yield, motivating the multi-fidelity split.","marker":"[9-11]"},{"why":"Provides the ignition metric chi_no_alpha that forms the ignition branch of the design objective.","marker":"[15,16]"},{"why":"Supplies the burn-fraction versus areal-density relation used for the burn-propagation branch of the objective.","marker":"[17]"},{"why":"Gives the hydro-equivalent scaling rules used to extrapolate 25 kJ designs to 2 MJ.","marker":"[19]"},{"why":"Provides the pulse-shape parameterisation and a reference automated Bayesian-optimisation approach for direct-drive designs.","marker":"[21]"},{"why":"Describes the Monte Carlo burn package in the simulation code used for the 2 MJ burn-on calculations.","marker":"[25]"},{"why":"Underpins the deep-ensemble method that gives the neural-network surrogates their uncertainty estimates.","marker":"[42]"},{"why":"Introduces transfer learning for inertial fusion models, the mechanism for carrying 1D structure into the 2D surrogate.","marker":"[47]"},{"why":"Defines the log Expected Improvement acquisition function used in the Bayesian optimisation loop.","marker":"[51]"}],"fun_headline_variants":["AI-driven design yields stable fusion target, 217x burn","Cheap 1D runs guide AI to stable laser fusion design","Multifidelity AI finds 25 kJ target that scales to 2 MJ","Stability-focused AI design boosts fusion burn 217-fold","Neural surrogates pick robust implosion design for fusion"],"cache_read_input_tokens":23808,"weakest_assumption_plain":"The 2 MJ high-gain numbers rest on the assumption that a hydrodynamically equivalent implosion can be produced up to stagnation by simple scaling, without re-tuning the laser pulse and target; if that fails, the scaled yields are extrapolations rather than confirmations.","fun_headline_variants_meta":{"raw":{"variants":["AI-driven design yields stable fusion target, 217x burn","Cheap 1D runs guide AI to stable laser fusion design","Multifidelity AI finds 25 kJ target that scales to 2 MJ","Stability-focused AI design boosts fusion burn 217-fold","Neural surrogates pick robust implosion design for fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000311,"raw_usage":{"total_tokens":1597,"prompt_tokens":722,"completion_tokens":875,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":786}},"tokens_in":466,"tokens_out":875,"duration_ms":9439,"temperature":1.0,"reasoning_tokens":786,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:44:43.154173+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the O2D design at 2 MJ with a properly hydro-equivalent re-tuned pulse and target, for example exchanging ablator mass for DT ice as hydro-equivalent ignition theory prescribes, and compare the burn-on yield against the reported 217x amplification; also fire the O2D target on a 25 kJ laser with seeded surface perturbations and compare yield and areal density with the 2D predictions. Agreement would confirm the scaled result, while a collapse toward the 1D-optimised design's 17x amplification would falsify it.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the burn-fraction versus areal-density relation used for the burn-propagation branch of the objective."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the hydro-equivalent scaling rules used to extrapolate 25 kJ designs to 2 MJ."},{"cited_title":"Automated and highly parallelized bayesian optimization scheme for direct drive fusion experiments on omega","cited_arxiv_id":null,"evidence_quote":"Provides the pulse-shape parameterisation and a reference automated Bayesian-optimisation approach for direct-drive designs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the Monte Carlo burn package in the simulation code used for the 2 MJ burn-on calculations."},{"cited_title":"Transfer learning to model inertial confinement fusion experiments","cited_arxiv_id":null,"evidence_quote":"Introduces transfer learning for inertial fusion models, the mechanism for carrying 1D structure into the 2D surrogate."},{"cited_title":"Unexpected improvements to expected improvement for bayesian optimization","cited_arxiv_id":null,"evidence_quote":"Defines the log Expected Improvement acquisition function used in the Bayesian optimisation loop."}],"review_version":1}