A self-adaptive Gaussian process regression approach for temperature fluctuation predictions in T-junction pipes of sodium-cooled fast reactors

Accurate determination of inner wall temperature fluctuations is critical for thermal fatigue assessment in sodium-cooled fast reactors (SFR) piping systems, but remains inaccessible for direct measurement due to extreme operational conditions involving high temperature and chemical activity of liquid sodium. To overcome this challenge, this study proposes a self-adaptive Gaussian process regression (GPR) approach. The large eddy simulations (LES) of hot and cold liquid sodium mixing in T-junction pipes are conducted to quantify intense thermal-fluid interactions, revealing that inner wall temperature fluctuations are significantly higher than those at the outer walls. Building on these insights, we develop a self-adaptive GPR approach that integrates tree-structured composite kernel optimization with gradient-based hyperparameter tuning. The resulting approach accurately predicts inner wall temperature fluctuations using only outer wall measurements and corresponding operational parameters, achieving a predictive performance of determination coefficient R2​>0.95, and retaining robustness (R2​>0.75) even when trained on limited datasets. The proposed self-adaptive GPR approach offers non-intrusive, real-time thermal diagnostics for SFR piping systems, utilizing composite kernels that afford clear physical interpretability. Moreover, it provides a promising tool for safety monitoring in reactor cores, heat exchangers, and other nuclear components requiring high-fidelity thermal transient analysis.