核事故发生后，应迅速、科学地进行事故后果评价，为后续的核应急措施提供科学的决策依据。核事故源项估算是核事故后果评价及核应急响应的关键环节。核事故源项反演，是基于监测数据及大气扩散模型，利用数值方法对事故释放源项信息进行估算的过程。传统的源项反演方法计算量大，计算成本高，可反演的信息有限。结合近年来迅速发展的机器学习算法，可有效提升反演计算效率，丰富可反演的有效信息。本研究将机器学习算法中经典的梯度升压决策树算法与贝叶斯反演方法结合，以浙江三门核电厂址为具体应用场景，使用拉丁超立方抽样等数值方法有效地对核事故源项释放率、有效释放高度、气象条件等信息进行反演。为解决传统反演方法计算成本较高的问题，本文利用机器学习模型替代扩散模型进行浓度计算。本研究中选取了源项释放率、有效释放高度、监测点极坐标信息等变量作为可变输入变量，基于前人研究经验，利用考虑复杂地形和建筑效应的CALMET（California Meteorological Model，CALMET）风场模型与RIMPUFF（Ris? Mesoscale Puff Model，RIMPUFF）大气扩散模型耦合而建立的CALMET-RIMPUFF扩散模型进行大量模拟计算，生成用于训练机器学习模型的模拟数据集。在机器学习模型的训练环节，分别选用集成学习算法中的梯度升压决策树算法（Gradient Boosting Decision Tree，GBDT）和反向传播神经网络算法（Back Propagation Neural Network，BPNN）对数据集进行训练，以生成可快速、准确预测固定监测点处浓度值的机器学习模型。同时，对影响两种模型拟合效果的若干重要超参数进行了调试研究和重要指标的敏感性分析。结果表明，初始超参数设置条件下的两种机器模型均可较准确地完成预测任务；推荐超参数设置条件下，GBDT模型的R2-score指标（R2决定系数评分指标）可达0.99以上，平均绝对误差与均方根误差均小于0.2，且其计算代价小于BPNN模型。因此，可将推荐超参数设置条件下的GBDT模型运用到反演计算中，替代CALMET-RIMPUFF扩散模型进行高效的数值计算。最后，本研究结合贝叶斯反演方法与GBDT算法，建立了基于机器学习的核事故源项反演方法。通过约束模拟结果与监测结果的相关性系数和均方误差，对三门风洞实验中的源项信息和部分气象信息的最大似然解进行了求解。结果显示，反演结果中的源项释放率、有效释放高度、气象条件等与实验设置相近。

After the occurrence of a nuclear accident, the consequences of the accident should be evaluated quickly and scientifically to provide a scientific decision-making basis for subsequent nuclear emergency measures. The estimation of the source term of a nuclear accident is a key link in the evaluation of nuclear accident consequences and nuclear emergency response. The inversion of the source term of a nuclear accident is a process of using numerical methods to estimate the source term information of the accident release based on monitoring data and atmospheric diffusion models. The traditional source term inversion method has a large amount of calculation and high calculation cost, and the information that can be inverted is limited. Combined with the rapid development of machine learning algorithms in recent years, it can effectively improve the efficiency of inversion calculations and enrich the effective information that can be inverted. This study combines the classic Gradient Boosting Decision Tree (GBDT) algorithm in machine learning algorithms with Bayesian inversion methods, takes the Zhejiang Sanmen nuclear power plant site as a specific application scenario, and uses numerical methods such as Latin hypercube sampling to effectively analyze the source of nuclear accidents. The release rate, effective release altitude, meteorological conditions and other information are inverted.In order to solve the problem of high computational cost of traditional inversion methods, this paper uses a machine learning model to replace the diffusion model for concentration calculation. In this study, variables such as the source term release rate, effective release height, and polar coordinate information of the monitoring point were selected as variable input variables. Based on previous research experience, the CALMET (California Meteorological Model, CALMET) wind field model that considers the complex terrain and architectural effects is coupled with the RIMPUFF (Ris? Mesoscale Puff Model, RIMPUFF) atmospheric diffusion model, and the obtained CALMET-RIMPUFF diffusion model is used for a large number of simulations calculate. The simulated data will be used to train machine learning modelsIn the training link of the machine learning model, the Gradient Boosting Decision Tree algorithm and the Back Propagation Neural Network algorithm (BPNN) in the integrated learning algorithm are used to train the data set. To generate a machine learning model that can quickly and accurately predict the concentration value at a fixed monitoring point. At the same time, some important hyperparameters that affect the fitting effect of the two models are debugged and analyzed and the sensitivity of important indicators is analyzed. The results show that the two machine models under the initial parameter setting conditions can complete the prediction task more accurately; under the recommended hyperparameter setting conditions, the R2-score index (R2 determination coefficient score index) of the GBDT model can reach more than 0.99. Both the average absolute the error and the root mean square error are less than 0.2, and the cost time of the calculation is less than the BPNN model. Therefore, the GBDT model under the recommended hyperparameter settings can be used in the inversion calculation, instead of the CALMET-RIMPUFF diffusion model for efficient numerical calculation.Finally, this research combines Bayesian inversion method and GBDT algorithm to establish a machine learning-based inversion model of nuclear accident source terms. By constraining the correlation coefficient and mean square error between the simulation results and the monitoring results, the maximum likelihood solution of the source term information and some meteorological information in the Sanmen wind tunnel experiment was solved. The results show that the source term release rate, effective release height, and meteorological conditions in the inversion results are similar to the experimental settings.