Predicting Global Heat Flow Using Machine Learning

Geothermal heat flow (GHF) is critical for constraining the thermal structure of the lithosphere and understanding geodynamic processes, yet global GHF measurements remain spatially insufficient. We apply three machine learning approaches to predict global GHF combined with geological and geophysical predictors that describe crustal and lithospheric structure. Sobol sensitivity analysis identifies Moho depth, crustal thickness, and uppermost asthenospheric mantle P-wave velocity as dominant controls, corresponding to the steady state crustal radiogenic and mantle conductive components of the three component heat flow model. Regional validation in Australia and the Tibetan Plateau confirms close agreement with measurements and prior models. The predictions reproduce the expected tectonic heat flow hierarchy, with low values over stable cratons and elevated values across active orogens, rifts, and subduction zones, marking collisional, shear, and extensional processes. The continuous and quantitatively constrained heat flow maps provide thermal boundary conditions for lithospheric modeling, geothermal assessment, and basin hydrocarbon maturation.