高级检索
黄美容, 黄永顺, 王德鑫, 张苏雅拉吐. 中子诱发裂变反应截面的贝叶斯神经网络预测模型[J]. 原子核物理评论, 2024, 41(1): 396-401. DOI: 10.11804/NuclPhysRev.41.2023CNPC85
引用本文: 黄美容, 黄永顺, 王德鑫, 张苏雅拉吐. 中子诱发裂变反应截面的贝叶斯神经网络预测模型[J]. 原子核物理评论, 2024, 41(1): 396-401. DOI: 10.11804/NuclPhysRev.41.2023CNPC85
Meirong HUANG, Yongshun HUANG, Dexin WANG, Suyalatu ZHANG. Bayesian Neural Network Prediction Models for Neutron-induced Fission Reaction Cross Sections[J]. Nuclear Physics Review, 2024, 41(1): 396-401. DOI: 10.11804/NuclPhysRev.41.2023CNPC85
Citation: Meirong HUANG, Yongshun HUANG, Dexin WANG, Suyalatu ZHANG. Bayesian Neural Network Prediction Models for Neutron-induced Fission Reaction Cross Sections[J]. Nuclear Physics Review, 2024, 41(1): 396-401. DOI: 10.11804/NuclPhysRev.41.2023CNPC85

中子诱发裂变反应截面的贝叶斯神经网络预测模型

Bayesian Neural Network Prediction Models for Neutron-induced Fission Reaction Cross Sections

  • 摘要: 核数据是核物理基础研究与核工程及核技术应用的基础数据。中子诱发核裂变反应截面是核数据的主要组成部分,在先进核能系统研发及核天体物理研究中具有重要应用。本工作基于贝叶斯理论,采用4层隐藏层每层200个神经元的前馈神经网络,利用马尔科夫链蒙特卡罗法(Markov Chain Monte Carlo simulation method, MCMC)和KL散度(Kullback–Leibler divergence, KL),针对238U、232Th和239Pu中子诱发裂变核反应截面数据,建立了以入射能量和裂变截面作为输入参数和输出参数的贝叶斯神经网络计算模型。选取中子能区为1~200 MeV的238U、232Th和1~100 MeV的239Pu裂变截面实验数据和评价数据进行模型训练,利用MCMC方法构建马尔科夫链来近似目标分布,结合KL散度约束方法对损失函数进行优化。研究结果表明,贝叶斯神经网络预测模型能够很好地再现1~200 MeV的238U、232Th和1~100 MeV的239Pu裂变截面数据,并对更高能量区间进行合理预测,为解决目前入射中子能量大于200 MeV的 238U、232Th和大于100 MeV的239Pu实验数据存在严重分歧问题及核数据评价提供重要参考。

     

    Abstract: Nuclear data is the basis for nuclear physics fundamental research, nuclear engineering, and applications in nuclear technology. Neutron-induced nuclear fission reaction cross-sections are a primary component of nuclear data, with significant applications in advanced nuclear energy system development and nuclear astrophysics research. This work, based on Bayesian theory, employs a feedforward neural network with four hidden layers. The Markov Chain Monte Carlo (MCMC) simulation method and Kullback–Leibler divergence (KL) are utilized. A Bayesian neural network computational model is established for neutron-induced fission reaction cross-section data of 238U, 232Th, and 239Pu, with incident energy and fission cross-section as input and output parameters, respectively. Experimental data and evaluation data of fission cross-sections for 238U and 232Th in the neutron energy range of 1~200 MeV and for 239Pu in the range of 1~100 MeV are selected for model training. The MCMC method is employed to construct a Markov chain to approximate the target distribution, and the KL divergence constraint method is combined to optimize the loss function. The research results indicate that the Bayesian neural network prediction model can effectively reproduce experimental data from literature, demonstrating strong predictive capabilities. This provides a reference for addressing the significant discrepancies in experimental data for 238U and 232Th with incident neutron energies greater than 200 MeV and for 239Pu with energies greater than 100 MeV, as well as for nuclear data evaluation.

     

/

返回文章
返回