An interpretable LSTM deep learning model predicts the time-dependent swelling behavior in CERCER composite fuels

Abstract

The study proposes a reliable and trustworthy data-driven approach with high precision and low computational cost, to predict the multiscale irradiation swelling behavior within CERCER composite fuels. A Long Short-Term Memory (LSTM) deep learning model is utilized to process the sequential data obtained from high-fidelity multiscale simulations. The model is evaluated using the metrics of accuracy, efficiency, interpretability, and generalizability. Compared to the feedforward and recurrent neural network models, the results indicate that the LSTM is superior at capturing the time-dependent character of the swelling progression and accurately catching the initiation of the critical recrystallization phenomena. The global and local interpretations via Shapley additive explanation theory (SHAP) agree with the mechanism analysis in simulations and experimental findings. The former identifies that the loading parameters of fission rate and density are the primary determinants in modulating the swelling deformation, while the latter uncovers the competing mechanism between temperature and hydrostatic pressure. The proposed model showcases its capacity to generalize on a time-consuming simulation involving a time-discrete algorithm. It guarantees a rapid swelling prediction in 36.4 secs that greatly reduces the time needed for 20.7 h in FE simulations.