Optimiser les essais d’irradiation du thorium grâce à une IA bayésienne

Hypothèse formelle

If the scheduling of nuclear irradiation experiments is optimized using a sequential Bayesian experimental design (SBED) framework with a multi-fidelity Gaussian process (GP) surrogate and explicit model discrepancy correction, then the posterior variance of the 233U production rate estimate will be reduced by at least 20% compared to a static design, because the acquisition function will preferentially select irradiation conditions that maximize information gain about the high-fidelity model parameters, while the discrepancy term will absorb and quantify the systematic bias between the low-fidelity simulator (WIMS-AECL) and the physical reactor.

Titre donné par l’affûteur : Sequential Bayesian Experimental Design for Calibration of Thorium Fuel Irradiation Simulations: A Methodological Framework for 233U Production Estimation

Contre-preuves

1. This paper is entirely unrelated to nuclear engineering or experimental design, indicating a potential mismatch in the literature search results.

   Gravité mineureNitrate isotopic composition of sequential Hurricane Harvey wet deposition: Low latitude NOx sources and oxidation chemistry

2. While supporting surrogate modeling in engineering, this paper focuses on pipeline transport, not nuclear irradiation, and does not address experimental design or Bayesian methods.

   Gravité mineureDeep-learning-based surrogate model for fast and accurate simulation in pipeline transport

3. This paper applies information gain in a machine learning context (search agents) and is not directly relevant to nuclear engineering or experimental design.

   Gravité mineureInformation Gain-based Policy Optimization: A Simple and Effective Approach for Multi-Turn Search Agents (2025)

Inconnues et conditions de validité

Inconnues identifiées

• The exact functional form and magnitude of the model discrepancy between WIMS-AECL and the physical reactor are unknown and must be estimated from the data.
• The optimal number of sequential steps (3 vs. 4 vs. 5) to balance information gain against cumulative cost is unknown and will be a subject of the study.
• The sensitivity of the optimal design to the specification of the cost function (e.g., linear vs. non-linear) is unknown.

Conditions de validité

• Neutron flux must remain below 1e14 n/cm²/s.Justification : Above this threshold, the assumption of linearity in the burnup reaction rate breaks down, and the GP surrogate’s smoothness assumption may be violated due to non-linear feedback effects (e.g., temperature reactivity coefficients).
• The fuel composition must remain in the thoria-rich range (Thoria fraction > 0.90).Justification : The WIMS-AECL simulator has been validated for this composition range. Outside this range, the model discrepancy term may become non-stationary, violating the GP assumption.
• The total irradiation time must be less than 720 hours.Justification : Beyond this time, the accumulation of fission products and changes in fuel microstructure (e.g., swelling) introduce additional sources of uncertainty not captured by the proposed discrepancy model.

Mécanisme proposé

Chaîne causale

1. Step 1: A low-fidelity simulator (WIMS-AECL) is used to generate an initial training dataset of 20 simulation runs across the input space (irradiation time, flux, composition). The output is the predicted isotopic composition vector, including 233U concentration.
2. Step 2: A multi-fidelity Gaussian process (GP) surrogate model is trained on this data. The GP is augmented with a model discrepancy term, δ(x), following the Kennedy-O’Hagan framework, which is a second GP that captures the systematic bias between the simulator and the true physical process.
3. Step 3: An acquisition function (e.g., Expected Information Gain or a variant of Expected Improvement) is defined. This function quantifies the expected reduction in posterior variance of the 233U production rate parameter, given a candidate irradiation condition and measurement time. The acquisition function incorporates both the GP predictive variance and the variance of the discrepancy term.
4. Step 4: The acquisition function is optimized to select the next irradiation condition (e.g., a specific flux and time point) that maximizes the expected information gain. This condition is then tested in a single, targeted irradiation experiment.
5. Step 5: The experimental measurement of 233U concentration (via mass spectrometry) is used to update the GP surrogate and the discrepancy term via Bayesian inference. The posterior distribution of the 233U production rate is refined.
6. Step 6: Steps 3-5 are repeated sequentially for a total of 3-4 experimental conditions, with the acquisition function guiding the selection of each new condition based on the accumulated knowledge, thereby minimizing the number of expensive experiments needed to achieve a target precision.

Postulats clés

• The low-fidelity simulator (WIMS-AECL) is a smooth and continuous function of the input parameters, allowing for accurate GP emulation.
• The model discrepancy term δ(x) is a stationary Gaussian process, meaning its correlation structure is translation-invariant.
• The measurement error of 233U concentration via mass spectrometry is Gaussian with a known, constant variance.
• The true physical process is a deterministic function of the input parameters (irradiation time, flux, composition), with all variability arising from measurement error and model discrepancy.
• The cost of each irradiation experiment is dominated by reactor time and post-irradiation analysis, and is independent of the specific input conditions.

Cadre théorique

Bayesian Optimal Experimental Design (BOED) and Kennedy-O’Hagan (KOH) model calibration framework, applied within the context of nuclear engineering simulation and experiment.

Variables

Prédictions falsifiables

1. 1

   The SBED framework will reduce the posterior variance of the 233U production rate by at least 20% compared to a static design (e.g., a space-filling Latin Hypercube design) after 4 experimental runs.

   Borne quantitative

   The ratio of posterior standard deviations (Static/SBED) will be ≥ 1.25, corresponding to a ≥20% reduction in variance.

   Méthode de mesure

   Posterior variance is computed from the GP surrogate after 4 sequential experiments (SBED) and after 4 static experiments (LHS). The 233U concentration is measured via High-Resolution Inductively Coupled Plasma Mass Spectrometry (HR-ICP-MS) with a measurement uncertainty of ±0.5%.Test statistique A one-sided F-test for equality of variances (ratio of variances), with α = 0.05 and power = 0.80. The sample size is determined by the number of posterior samples (n=1000) drawn from the GP, ensuring the test is well-powered.

   Hypothèse nulle

   H0: The posterior variance of the 233U production rate after SBED is ≥ 80% of the posterior variance after a static design (i.e., no significant improvement).

2. 2

   The estimated model discrepancy term δ(x) will be non-zero and statistically significant at the 95% confidence level for at least one input condition.

   Borne quantitative

   The 95% credible interval of δ(x) will exclude zero for at least one point in the input space.

   Méthode de mesure

   The discrepancy term is estimated from the difference between the GP surrogate predictions and the experimental measurements of 233U concentration at the selected irradiation conditions.Test statistique A Bayesian hypothesis test based on the posterior distribution of δ(x). The null is rejected if the 95% highest posterior density (HPD) interval of δ(x) excludes zero for any input condition.

   Hypothèse nulle

   H0: The model discrepancy term δ(x) = 0 for all input conditions (i.e., the simulator is unbiased).

3. 3

   The SBED framework will require fewer experimental runs (N) to achieve a target posterior variance (e.g., 10% of the prior variance) compared to a static design.

   Borne quantitative

   The ratio of required runs (N_static / N_SBED) will be ≥ 1.5.

   Méthode de mesure

   Simulation study: The SBED and static design are simulated 100 times each using the multi-fidelity GP with a known discrepancy function. The number of runs needed to reach the target variance is recorded for each simulation.Test statistique A one-sample t-test on the log-transformed ratio of runs, with α = 0.05 and power = 0.80. The sample size (100 simulations) is justified by a priori power analysis assuming a standard deviation of 0.3 on the log-ratio.

   Hypothèse nulle

   H0: The ratio of required runs (N_static / N_SBED) ≤ 1.5.

Protocole expérimental

Références

9 références, toutes issues de Semantic Scholar. Une référence vérifiée est un article qui existe et qui est indexé par Semantic Scholar. Cela ne veut pas dire qu’il confirme l’idée.

1. 1
   O. Surer, M. Plumlee, Stefan M. Wild (2023). Sequential Bayesian Experimental Design for Calibration of Expensive Simulation Models.appui direct · 20 citations · doi:10.1080/00401706.2023.2246157Ce qu’en retient le documentaliste A sequential framework with a novel criterion for parameter selection that targets learning the posterior density of parameters improves the efficiency of calibration for expensive simulation models.Pertinence Directly supports the core mechanism of using sequential Bayesian experimental design to calibrate expensive simulation models, which is the proposed approach for optimizing irradiation experiments.
2. 2
   Xun Huan, Jayanth Jagalur, Youssef Marzouk (2024). Optimal experimental design: Formulations and computations.appui direct · 119 citations · doi:10.1017/S0962492924000023Ce qu’en retient le documentaliste A systematic survey of modern OED, from classical design theory to current research involving OED for complex models, establishing the methodological basis for the hypothesis.Pertinence Provides the theoretical foundation for optimal experimental design, which is the statistical framework proposed in the hypothesis for determining optimal irradiation conditions and measurement times.
3. 3
   Adam Foster, Desi R. Ivanova, Ilyas Malik et al. (2021). Deep Adaptive Design: Amortizing Sequential Bayesian Experimental Design.appui indirect · 137 citations · sans DOICe qu’en retient le documentaliste Deep Adaptive Design (DAD) amortizes the cost of adaptive Bayesian experimental design, allowing experiments to be run in real-time, which is crucial for practical implementation.Pertinence Supports the feasibility of real-time adaptive experimental design, which is relevant for making sequential decisions during irradiation campaigns where computational time may be limited.Fiche Semantic Scholar
4. 4
   B. Khuwaileh, P. Turinsky (2017). Surrogate based model calibration for pressurized water reactor physics calculations.appui direct · 18 citations · doi:10.1016/J.NET.2017.08.007Ce qu’en retient le documentaliste A scalable surrogate-based algorithm for model calibration in nuclear engineering reduces uncertainty in key reactor attributes based on experimental and operational data.Pertinence Demonstrates the use of surrogate models for calibration in nuclear engineering, validating the feasibility of the emulation approach proposed in the hypothesis for reactor physics models.
5. 5
   Mandana Kariminejad, D.P. Tormey, Caitriona M. Ryan et al. (2024). Single and multi-objective real-time optimisation of an industrial injection moulding process via a Bayesian adaptive design of experiment approach.appui par analogie · 22 citations · doi:10.1038/s41598-024-80405-2Ce qu’en retient le documentaliste Bayesian adaptive design of experiments (ADoE) is an iterative process where previous experiment results inform the selection of subsequent experiments, reducing the number of experiments needed.Pertinence Provides an analogous application of Bayesian adaptive design of experiments in an industrial engineering context, demonstrating the practical utility of the proposed methodology in a different domain.
6. 6
   Florian Gisperg, Robert Klausser, M. Elshazly et al. (2025). Bayesian Optimization in Bioprocess Engineering—Where Do We Stand Today?.appui par analogie · 36 citations · doi:10.1002/bit.28960Ce qu’en retient le documentaliste Bayesian optimization combines machine learning with decision-making to optimally utilize information gained during experimentation, balancing exploration and exploitation.Pertinence Shows the successful application of Bayesian optimization in another engineering field (bioprocess), supporting the transferability of the proposed methodology to nuclear engineering.
7. 7
   Dario Azzimonti, D. Ginsbourger, C. Chevalier et al. (2016). Adaptive Design of Experiments for Conservative Estimation of Excursion Sets.appui indirect · 50 citations · doi:10.1080/00401706.2019.1693427Ce qu’en retient le documentaliste The proposed strategy of minimizing false negatives in conservative estimation achieves competitive performance using Gaussian process models for expensive-to-evaluate functions.Pertinence Supports the use of Gaussian process emulation and adaptive design for expensive-to-evaluate functions, which is the statistical machinery proposed for the reactor physics simulator.
8. 8
   Sayan Ghosh, Jesper Kristensen, Yiming Zhang et al. (2019). A Strategy for Adaptive Sampling of Multi-fidelity Gaussian Process to Reduce Predictive Uncertainty.appui indirect · 19 citations · doi:10.1115/detc2019-98418Ce qu’en retient le documentaliste Multi-fidelity Gaussian process modeling with adaptive sampling reduces predictive uncertainty by partitioning uncertainty based on fidelity level and associated cost.Pertinence Supports the use of multi-fidelity Gaussian processes for surrogate modeling, which could be relevant for combining low-fidelity simulations with high-fidelity experiments in irradiation campaigns.
9. 9
   Raghav Gnanasambandam, Bo Shen, A. C. Law et al. (2024). Deep Gaussian process for enhanced Bayesian optimization and its application in additive manufacturing.appui par analogie · 14 citations · doi:10.1080/24725854.2024.2312905Ce qu’en retient le documentaliste Bayesian Optimization (BO) is a methodology to optimize black-box functions where the input-output relationship is unknown and expensive to compute, applicable to engineering design.Pertinence Demonstrates the application of Bayesian optimization with Gaussian processes in another manufacturing domain, supporting the generalizability of the approach to nuclear experiment planning.

Nouveauté

Score de nouveauté : 0,70 sur 1 · Verdict : jugée nouvelle

Ce score est attribué par un agent à partir des travaux trouvés. C’est une estimation, pas une mesure. Comment ce score est produit

Travaux les plus proches

• Sequential Bayesian Experimental Design for Calibration of Expensive Simulation Models (2023)Différence avec l’idée : This paper applies sequential Bayesian experimental design to calibrate expensive simulation models in general, but does not specifically address nuclear irradiation experiments or thorium-based fuel cycles. The hypothesis extends this methodology to a specific nuclear engineering application.
• Surrogate based model calibration for pressurized water reactor physics calculations (2017)Différence avec l’idée : This paper uses surrogate models for calibration in nuclear engineering but does not employ Bayesian experimental design for adaptive scheduling of experiments. The hypothesis proposes a more formal statistical framework for experiment planning.
• Applications of emulation and Bayesian methods in heavy-ion physics (2023)Différence avec l’idée : This paper demonstrates Bayesian emulation in nuclear physics but focuses on heavy-ion collisions rather than irradiation experiments for fuel production. The hypothesis targets a different nuclear engineering domain.

Lacunes et données

Lacunes identifiées

• No specific public dataset of nuclear irradiation experiments with sufficient metadata for applying Bayesian optimal design was identified in the provided papers.
• The literature does not provide specific runtime data for full-core or pin-cell reactor physics simulations, which is needed to estimate the computational cost of the proposed approach.
• No paper in the list directly addresses the application of Bayesian experimental design to thorium-based fuel irradiation experiments, leaving a specific application gap.

Données disponibles

• No specific dataset of nuclear irradiation experiments was identified in the provided papers. The papers provide methodological frameworks but not experimental datasets for thorium-based fuels.

Synthèse du panel

Note de consensus : 6,20/10 Moyenne des cinq notes, pondérée par la confiance que chaque relecteur déclare.

Verdict du méta-relecteur : publier

Points d’accord

• The Kennedy-O’Hagan (KOH) framework with an explicit discrepancy term is unanimously recognised as a methodologically sound and appropriate approach for addressing the systematic bias between the low-fidelity simulator (WIMS-AECL) and physical reality.
• The three-phase validation structure (in silico, minimal physical, full physical) with quantitative GO/NO-GO criteria is regarded as an example of progressive de-risking and scientific rigour, although the statistical power of the early phases (n=4) is questioned.
• Prediction 3 (comparison of the number of trials required to reach a target variance) is identified as a more robust and more physically meaningful evaluation metric than the simple reduction of variance at a fixed number of trials, as it reflects the true cost of the experiments.

Points de désaccord

• The quantitative threshold of a 20% reduction in variance is judged arbitrary and not theoretically justified by the Domain expert, whereas the Methodologist regards it as a clear decision criterion but statistically fragile with n=4.
• The Contrarian asserts that the measured reduction in variance will be an artefact of the method (a reduction in the variance of the surrogate GP, not of the true uncertainty in the production rate), whereas the Domain expert and the Methodologist consider that the KOH framework with discrepancy is able to capture this uncertainty correctly.
• The Contrarian insists on the absence of any account taken of fundamental nuclear uncertainties (correlated and non-Gaussian cross-sections) and of the non-stationarity of the bias, whereas the Domain expert recommends that these points be treated as necessary but non-blocking improvements.
• The Funding strategist judges the TRL (1-2) too low for European collaborative calls, whereas the Domain expert and the Methodologist consider that in silico validation is a sufficient preliminary step for a methodological research project.

Chemin critique

The most determining factor is the capacity of the framework to demonstrate, on simulated or historical data, that the separation between the simulator bias (discrepancy) and the measurement error is identifiable with a realistic number of experimental points (N=3-4), and that the measured variance reduction reflects a genuine improvement in the estimation of the 233U production rate, and not an artefact of the method. The success of this demonstration conditions the validity of the hypothesis and adoption by industrial partners and regulators.

Recommandation finale

The panel recommends publication of the brief with a moderate level of confidence. The methodological framework (SBED + KOH) is theoretically sound and innovative, and the three-phase validation structure is a major asset. However, the Contrarian’s criticisms regarding the identifiability of the discrepancy with a small number of points and the absence of consideration of fundamental nuclear uncertainties must be addressed in Phase 1 (in silico) of the protocol. The brief should be published with the explicit mention that the in silico validation must include a sensitivity analysis to the prior on the hyperparameters of δ(x) and a demonstration of bias/error separation on simulated data. The Funding strategist’s recommendation to target exploratory funding (ANR JCJC) is validated as the primary funding pathway.