Thermal conductivity prediction of UO₂-BeO composite fuels and related decisive features discovery via convolutional neural network

Improved thermal conductivity (TC) in UO₂-based fuels is critically important for the security of nuclear reactors. Optimizing the microstructure of UO₂ pellets mixed with high thermally conducting materials such as BeO is an effective way to achieve this goal. In this paper, we investigate the relationship between the TC of the UO₂/BeO composite and their corresponding microstructures by using a convolutional neural network (CNN). The dataset is created by the finite element method. Using the deep transfer learning technique, a well-established CNN (ResNet101) successfully predicts the TC based on the highly similar microstructures while the lightweight CNN fails. Meanwhile, the microstructural features impacting the TC of the UO₂/BeO composite are determined by the ablation class activation map and subsequently verified. Moreover, a proposed microstructural feature contributes to the improved performance of the state-of-the-art machine learning model on our task, indicating its essential role in the TC of the UO₂/BeO composite. This work demonstrates that the critical microstructural features that significantly affect the TC can be discovered by deep learning-based methods and may provide a guideline for enhanced TC of UO₂/BeO fuel via careful design of its microstructure.