Abstract
Accurate determination of the geothermal gradient is critical for assessingthe geothermal energy potential of a given region. Of particular interest isthe case of Colombia, a country with abundant geothermal resources. A historyof active oil and gas exploration and production has left drilled boreholes indifferent geological settings, providing direct measurements of the geothermalgradient. Unfortunately, large regions of the country where geothermalresources might exist lack such measurements. Indirect geophysical measurementsare costly and difficult to perform at regional scales. Computational thermalmodels could be constructed, but they require very detailed knowledge of theunderlying geology and uniform sampling of subsurface temperatures to bewell-constrained. We present an alternative approach that leverages recentadvances in supervised machine learning and available direct measurements topredict the geothermal gradient in regions where only global-scale geophysicaldatasets and course geological knowledge are available. We find that a GradientBoosted Regression Tree algorithm yields optimal predictions and extensivelyvalidate the trained model. We show that predictions of our model are within12\% accuracy and that independent measurements performed by other authorsagree well with our model. Finnally, we present a geothermal gradient map forColombia that highlights regions where futher exploration and data collectionshould be performed.