Cross-scale Data-Driven Heat Flow Prediction in the Ordos Basin and Adjacent Areas: An XGBoost Model with SHAP Interpretability

Abstract

Summary Terrestrial heat flow is a fundamental parameter characterizing the thermal regime of sedimentary basins, but sparse and unevenly distributed measurements limit regional heat flow mapping. This study integrated cross-scale geological and geophysical features to compare XGBoost, Random Forest, and Deep Neural Network models in the Ordos Basin and adjacent areas in northern China, East Asia. XGBoost achieved the best overall predictive performance. The predicted heat flow ranged from 54.0 to 74.0 mW m⁻², with an average of 63.76 ± 3.88 mW m⁻², and generally increased from west to east within the Ordos Basin. Bootstrap analysis indicated a prediction uncertainty of 1.5–3.3 mW m⁻². Shapley Additive exPlanations analysis showed that Moho depth was the dominant contributing feature at the regional scale, whereas sedimentary thickness showed greater importance in the interior tectonic units of the Ordos Basin. The spatial variations in feature contributions were broadly consistent with the regional geological and thermal setting, supporting the geological plausibility of the model interpretation. These results demonstrate the applicability of machine learning to regional heat flow prediction and provide a geologically interpretable framework for areas with limited heat flow observations.