Continental-scale geothermal heat flow patterns in the United States revealed by a spatial correlation recurrent neural network

SUMMARY

Geothermal heat flow (GHF) provides a first-order constraint on lithospheric thermal structure, yet continental-scale prediction remains challenging due to sparse, uneven and strongly non-stationary observations. Most existing machine-learning approaches treat GHF measurements as spatially independent, limiting their ability to capture geological coherence across tectonic provinces. Here, we develop a spatial correlation recurrent neural network (SCRNN) that explicitly incorporates spatial dependencies among multiple variables to predict continental-scale GHF. By employing SCRNN, we produce a high-resolution GHF distribution of the contiguous United States. Higher GHF values are concentrated in tectonically active western regions characterized by crustal thinning, lithospheric extension and fault systems. Lower GHF values are in the geologically stable central and eastern Craton of North America. SHapley Additive exPlanations analysis indicates that lithosphere–asthenosphere boundary depth, Moho depth, Curie-point depth and fault systems exert first-order controls on GHF patterns. These results demonstrate that SCRNN captures meaningful controls on continental GHF.