UC Berkeley

The accuracy of neutrons modeling and simulation tools strongly depends on the quality of the nuclear data. Data libraries are generated by evaluators combining physics-based model codes and experimental data. There are many instances where experimental data are not available, are not reported rigorously or are discordant. In such cases, the evaluators need to make an expert judgment exposing the generated data to human bias and large uncertainties. This work proposes to support the evaluators’ complex tasks by leveraging Machine Learning (ML) and Artificial Intelligence (AI). Two proof-of-concept ML models, a Decision Tree and K-Nearest-Neighbor, were developed to fit nuclear data from the EXFOR database in order to infer neutron induce reaction cross sections. Both models were used to predict nuclear data for 233U, a well-characterized isotope in literature, and 35Cl, a less studied but important nuclide for some advanced nuclear reactors. The predicted values for 233U were validated using the 233U Jezebel benchmark in Serpent2 model. The predicted values for 35Cl(n,p) cross section were compared against recent new measurement not available in EXFOR. The predicted ML/AI values matched more accurately the new measurements than any of the evaluated data libraries, which overestimate experimental results by up to a factor of five. In turn, the proof-of-concept models explored in this work, reliant on learning underlying patterns of cross section data from other radionuclides, demonstrate evidence that ML models can aid traditional physics-guided models and have a role to play in nuclear data evaluations. Furthermore, incorporating ML models in the nuclear data pipeline can allow evaluators to make faster bias-free decisions in areas of uncertainty as well as better inform future data measurement campaigns on areas of greatest sensitivity in EXFOR.