A physically motivated feature engineering and proxy validation strategy for XGBoost-based reconstruction of thermal conductivity in supercritical hydrocarbon fluids

In advanced thermal management systems involving supercritical hydrocarbon fluids, accurate thermal-conductivity data are essential but remain scarce in high-temperature, transcritical, and supercritical regimes. This study develops a feature-based XGBoost framework to reconstruct thermal conductivity using physically motivated thermophysical descriptors and accessible low-temperature data. Motivated by the corresponding-states principle, the framework examines whether low-temperature data from related hydrocarbons can support plausible high-temperature reconstruction for RP-3 aviation kerosene. The framework is evaluated through two complementary validation stages: Leave-One-Substance-Out Cross-Validation (LOSOCV), in which one hydrocarbon is withheld during training, and anchored high-temperature extrapolation validation using proxy fluids, in which low-temperature liquid data are used as anchors while high-temperature transcritical/supercritical data remain unseen. Across the investigated hydrocarbon surrogates, XGBoost achieves an average MAPE of 4.27% in LOSOCV. Feature-ablation results show that adding μCp reduces the mean high-temperature extrapolation MAPE from 4.61% to 4.18%. The framework is further applied to RP-3 aviation kerosene, where the low-temperature anchored model reproduces the available low-temperature trend with a MAPE of 1.6%. In the high-temperature region, the reconstructed values show an average relative deviation of 16.6% from extended-corresponding-states-based auxiliary reference estimates. Because direct high-temperature RP-3 thermal-conductivity measurements are unavailable, the RP-3 results are interpreted as an engineering consistency assessment rather than direct experimental validation.