Machine learning approaches to equilibrium burnup analysis for Molten Salt Reactor

Abstract

During the operation of a liquid-fueled molten salt reactor (MSR), the deep depletion to an equilibrium state can be attained with online refueling and reprocessing techniques. However, the equilibrium burnup simulation of an MSR using current depletion codes requires considerable computational cost and time, and even fails when the equilibrium conditions are unsatisfied. To efficiently evaluate the specific characteristics of the equilibrium burnup in MSRs, a fast-filtering model for equilibrium state and a fast prediction model for equilibrium neutronic properties are developed based on the machine learning (ML) technique. A database of 4860 cases including the neutronic information of equilibrium and non-equilibrium of a fuel assembly with variable geometric parameters in an MSR is first generated, and then the neutronic properties of 232Th-233U, 232Th-EU (enriched uranium) and 232Th-Pu fuel cycles are evaluated. Based on the MSR database, fifteen different ML classification algorithms are implemented to establish the fast-filtering model for equilibrium burnup state. Moreover, a model for the fast prediction of equilibrium neutronic parameters of neutron flux (Φ) and breeding ratio (BR) is developed with fifteen ML regression algorithms. Considering the various performance metrics for measuring the predicted performances of ML models in the classification and regression, the LightGBM (LGBM) model is the most favorable for filtering the burnup state and predicting the neutronic parameters in MSRs. With the test data, the LGBM model can obtain the accuracy of about 0.98 for classification and the relative error of about 1.10% for regression.