Effective thermal conductivity prediction of dispersion nuclear fuel elements based on deep learning and property-oriented inverse design

Abstract

The in-pile thermal performance of dispersion nuclear fuel elements is crucial for reactor safety. The uncertainty of the thermal conduction throughout dispersion fuel is primarily influenced by the nonuniform distribution of fuel particles in the meat, especially the agglomeration behavior of fuel particles. In this paper, a new method has been developed for the rapid and accurate prediction of the effective thermal conductivity (ETC) of dispersion nuclear fuel elements based on the deep learning method. A deep learning model is trained to establish an implicit correlation between the microstructure of dispersion nuclear fuel elements and their ETC, enabling a predictive model to estimate ETC from two-dimensional microstructural diagrams. The dataset is generated using the finite element method, which takes into account the nonuniform distribution characteristic of fuel particles. The microstructures which affect the decision-making of the predictive model are demonstrated by the saliency map. Based on the predictive model, an inverse design method of the distribution of fuel particles was conducted for the specific ETC by metaheuristic algorithms. This study confirms the feasibility of directly evaluating ETC from microstructural images and provides a property-oriented inverse design method for the meat microstructure.