A physics-informed and transferable machine learning workflow for basin-scale heat flow prediction

Accurate prediction of heat flow is critical for geothermal resource assessment yet remains challenging in structurally complex basins because of sparse measurements and the coupled effects of multiple geological processes. We develop a physics-informed and transferable machine-learning workflow that integrates multi-source geological and geophysical parameters through a stacking ensemble model. Application to the Junggar Basin demonstrated improved basin-scale predictive performance relative to Gaussian-variogram Ordinary Kriging. Under the matched random-holdout evaluation, the stacking model achieved an RMSE of 3.2 mW/m 2 and an R 2 of 0.64, compared with 3.9 mW/m 2 and 0.48 for Gaussian Ordinary Kriging. Under repeated local spatial-block five-fold validation, the stacking model retained better performance, with an RMSE of 4.3 mW/m 2 and an R 2 of 0.38, compared with 4.4 mW/m 2 and 0.36 for Gaussian Ordinary Kriging. The interpretability analysis further differentiated the geological factors associated with heat-flow variation among tectonic units: sedimentary-cover thickness had the highest surrogate importance in the Central Depression (17%), lithospheric thickness in the Eastern Uplift (15%), and gravity and magnetic anomalies in the Western Uplift. These patterns support a hierarchical geological interpretation involving both shallow sedimentary structure and deeper crustal–lithospheric architecture. Application of the Junggar-trained model to the Tarim Basin reproduced the broad spatial heat-flow pattern, while localized discrepancies highlighted the need for basin-specific feature reassessment and calibration. These results establish an integrated, interpretable, and adaptable workflow for basin-scale heat-flow prediction and geological analysis in data-scarce sedimentary basins.