Skip to main content
QUICK REVIEW

[论文解读] Physics Informed Bayesian Machine Learning of Sparse and Imperfect Nuclear Data

Jiaming Liu, Yang Su|arXiv (Cornell University)|Feb 2, 2026
Nuclear reactor physics and engineering被引用 0
一句话总结

该论文将物理知识融入贝叶斯神经网络,以在稀疏数据下推断能量相关的独立裂变产额,使用基于GEF的先验和累计产额作为约束,在信息不足的学习下获得更准确且物理一致的结果。

ABSTRACT

The prevailing data-driven machine learning has been plagued by the absence of physics knowledge and the scarcity of data. We implement the physics-model informed prior into Bayesian machine learning to evaluate the energy dependence of independent fission product yields, which are crucial for advanced nuclear energy applications but only sparse and imperfect experimental data are available. The informative prior is the posterior after learning the generated data from fission models. Furthermore, cumulative fission yields are included as a physical constraint via a conversion matrix to provide augmented energy dependence. Our work demonstrated a truly Bayesian machine learning by incorporating comprehensive physics knowledges as a powerful tool to exploit the sparse but expensive nuclear data.

研究动机与目标

  • 由于数据稀缺且存在不完善性,说明需要在核数据中引入物理信息进行学习。
  • 开发一个贝叶斯神经网络框架,将物理先验融入以推断独立裂变产额的能量依赖性。
  • 使用将累计产额与独立产额相关的转换矩阵,通过物理约束增强能量依赖性。
  • 相对于无信息学习,证明在再现能量依赖性及裂变产额的细结构方面有改进。

提出的方法

  • 使用GEF模型为235U生成能量相关的独立产额作为物理数据D_phys用于先验构建。
  • 在对生成数据训练后,计算后验P(w1|D_phys),并将其作为后验P(w2)用于对实验数据D_expt的后续评估的先验。
  • 应用一个两隐藏层神经网络(每层22个神经元,tanh激活)来建模产额Y_i(A_i,Z_i,E_i)。
  • 通过MCMC执行贝叶斯推断,对网络权重进行采样并给出产额的95%置信区间。
  • 通过一个编码β衰变关系的转换矩阵,将独立产额与累计产额融合在一起,联合损失包含两种数据类型的卡方项。
  • 通过将Y_i^c(独立产额)与通过转换矩阵测得的累计产额t_j^e相联系,加入物理约束,使用全局噪声尺度sigma和数据驱动的加权。
Figure 1: Illustration of the physics model informed Bayesian machine learning for evaluations of independent fission yields, with and without physics constraints. The physics model generated data are used to train the informed priors for evaluation of measured data. The heterogenous cumulative yiel
Figure 1: Illustration of the physics model informed Bayesian machine learning for evaluations of independent fission yields, with and without physics constraints. The physics model generated data are used to train the informed priors for evaluation of measured data. The heterogenous cumulative yiel

实验结果

研究问题

  • RQ1物理信息先验如何改善稀疏裂变产额数据的贝叶斯学习?
  • RQ2将累计产额作为物理约束引入,是否能提升独立产额的能量依赖性与细结构?
  • RQ3物理信息先验对收敛性与准确性相对于无信息学习有何影响?
  • RQ4推断出的能量依赖性与已建立的评估(如JENDL-5、ENDF/B-VIII.1)在不同能量下的比较如何?

主要发现

  • 信息化学习在与评估数据的一致性方面显著优于无信息学习,归一化偏差从约5.3%降至约0.22%。
  • 物理先验使训练收敛速度比无信息学习更快,在对评估数据的损失值方面表现为更低。
  • 通过转换矩阵引入累计产额约束显著降低受限损失并更好地保证物理一致性,尽管存在数据不一致。
  • 信息化学习预测的能量依赖性和细结构(如奇偶振荡)与无信息结果不同,在3–14 MeV等多个能量点更符合物理预期。
  • 当使用物理信息先验时,该方法可实现非单调的能量依赖性和产额趋势的改进插值。
Figure 2: The inferences of neutron induced fission yields of 235 U at different incident energies. (a) mass yields with uninformed learning; (b) charge yields with uninformed learning; (c) mass yields with informed learning; (d) charge yields with informed learning. The evaluated JENDL data (black
Figure 2: The inferences of neutron induced fission yields of 235 U at different incident energies. (a) mass yields with uninformed learning; (b) charge yields with uninformed learning; (c) mass yields with informed learning; (d) charge yields with informed learning. The evaluated JENDL data (black

更好的研究,从现在开始

从论文设计到论文写作,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。