In-pile heat conduction model of the dispersion nuclear fuel plate with particle agglomeration. Part II: Predicting the effective thermal conductivity under the in-pile thermal transfer pattern based on a deep neural network

Abstract

In-pile dispersion nuclear fuel plate elements incorporate multiple internal heat sources and evolving interaction layers. The effective thermal conductivity of this complex heat conduction structure exhibits multivariate and highly nonlinear characteristics, requiring accurate prediction by combining numerical methods and machine learning techniques. This study developed an automated high-throughput workflow to rapidly simulate massive number of dispersion meat models with diverse microstructures, and further established the machine learning database for predicting the effective thermal conductivity. Influences of internal heat sources within fuel particles, the particle agglomeration and interaction layers were considered simultaneously in modeling. Through utilizing the database constructed by high-throughput computing, the deep neural network was successfully applied to accurately forecast the effective thermal conductivity of the in-pile dispersion meat. Furthermore, comparisons of the predictive performances determined by different distance-based point pattern analysis methods indicated that the radial distribution function (g(r)) served as the most effective approach for characterizing spatial discreteness in predicting the effective thermal conductivity. This work demonstrates that thermal conductivity of ceramic dispersion nuclear fuel plate elements can be enhanced via optimizing the critical microstructural features.