Yuanzineng kexue jishu (Apr 2023)
Learning Fast Neutron Cross Section by Deep Neural Network
Abstract
Nuclear data, especially neutron nuclear data, are the basic data for nuclear science and engineering application. Traditional nuclear data evaluation is time-consuming and labor-intensive, and is easily influenced by human factors. Machine learning technology is expected to enhance the ability of nuclear data evaluation. In order to explore the feasibility of using machine learning method to assist nuclear reaction data evaluation, a fully connected deep neural network algorithm was used to learn neutron cross section data to obtain the trained model, and the test data were used to examine the prediction ability of the model. In order to avoid the complexity of resonant region cross section, the neutron cross sections in the fast neutron region in the general evaluation nuclear library were used as the data set, with which the neural network model was trained, verified and tested. The total neutron cross sections and elastic scattering cross sections of 12 uranium isotopes from ENDF/B-Ⅷ.0 in fast neutron region were extracted. The cross sections of 230U were used as test data to be predicted, the cross sections of 232U were used as verification data, and the cross sections of other 10 nuclides were used as training data. In order to obtain a neural network model with predictive ability, a series of neural network models were trained by using training data, and then the best model was selected by using verification data to predict test data. In order to find a suitable network structure, neural network models were constructed with a certain step size (10 neurons) within the number of hidden layers 2-10 and the number of neurons in each layer less than 100. After completing all the model training, the best model was selected by using the validation data, the cross section data of test nuclides were predicted by using the best model, and the prediction ability was evaluated by comparing with the data of evaluation library. For the total cross section, taking (A, E) as the features, satisfactory prediction results can be obtained, however, for elastic scattering cross sections, it is difficult to get satisfactory prediction results by taking (A, E) as the features. Further increasing the neutron separation energy Sn of the target nucleus as feature, satisfactory results can be obtained. Validation and testing show that the trained neural network model can well grasp the variation law of cross sections with nuclides and incident neutron energy in the evaluation library, so the model has strong predictive ability for “unknown” neutron cross sections of a nuclide. Therefore, neural network algorithm is believed to have the potential to become a new way to evaluate nuclear data.
