Benchmarking Machine Learning and Seismic-Derived Heat Flow Models to Decipher the Lithospheric Thermal Architecture of the Arabian-Nubian Shield

<title>Abstract</title> <p>Geothermal energy is a cornerstone of the global transition to sustainable power; however, its exploration is frequently hindered by sparse heat flow data in tectonically complex regions. We present a novel benchmarking framework that compares two fundamentally distinct paradigms for terrestrial heat flow prediction: a data-driven Random Forest Regression (RFR) model and a physics-based Velocity-Temperature (V-T) conversion. Applied to the tectonically diverse Arabian-Nubian Shield (ANS), we integrated sixteen multi-parametric geological and geophysical observables to train the RFR model, while simultaneously deriving deep thermal structures from S-wave tomography at depths of 56–200 km. Our results quantify significant lateral and vertical thermal heterogeneities, identifying extreme mantle temperatures (&gt;1600 K) and lithospheric thinning (&lt;80 km) beneath the Red Sea and Afar rifts, contrasting with the deep, cold lithospheric roots (&gt;200 km) of the Arabian Platform. While both methods converge on consistent first-order thermal patterns (r = 0.86), validation against 4,240 measurements reveals that the RFR model achieves higher predictive precision (R² = 0.92) compared to the V-T approach (R² = 0.79). This discrepancy highlights their complementary nature: RFR effectively captures localized, non-linear crustal anomalies through multi-data integration, whereas the V-T method excels at characterizing large-scale lithospheric processes and providing physical constraints where surface data are absent. Feature importance analysis identifies Moho and LAB depths as the primary drivers of the region’s thermal architecture. This study demonstrates that this dual-methodology approach offers a robust, scalable framework for geothermal resource assessment in complex tectonic settings worldwide.</p>