Geothermal flow in Northern Morocco: A machine learning approach

Abstract

To model the geothermal flow in northern Morocco, we apply a machine learning (ML) approach by analyzing the geological and geophysical data at boreholes, where direct heat flow data is available. The data includes geothermal flow, geothermal gradient, Bouguer gravity, aero-magnetic field, tectonic fracture density and the proximity level to the major faults, thermal spring data (location, temperature, density), geochronological age of formations, Earthquake data (density, magnitude, location), and the proximity level to the recent volcanism in Morocco. These data represent factors that generally control the spatial variation of geothermal flow. For this purpose, we used four ML techniques: Multi-Layer Perceptron, Support Vector Regression, K-Nearest Neighbors, and Decision trees, and examined the impact of parameter settings for each ML technique on the performance results. In specifically, we examine two parameter tuning methods: Grid Search (GS) and the Python Tool's default parameters. To determine the significance of the performance differences and rank ML techniques according to their performances, the Skott-Knott test, and the Borda Count voting system were investigated. We identified the optimized MLP by means of GS as the best ML technique. The cartographic representation of predicted geothermal flow values, by the optimized MLP (GS + MLP) model over the northern Morocco, shows areas of high geothermal flow values. These areas often correspond to magmatic intrusions at depth connected to the regional geodynamic context. The lowest values of geothermal flow are less than 40 mW m−2 and are mainly predicted in the Anti-Atlas chain of Precambrian age and in which is part of the West African Craton. This method is a new approach that may help to identify areas with high geothermal potential based on geological and geophysical data.