Predicting Geothermal Suitability For Power Production At Global Scale Through Ai

Today, many scientific disciplines, including Earth Science, are increasingly using Artificial Intelligence (AI) methodologies, particularly those based on Machine Learning (ML) models. Geothermal exploration and resource assessment has recently started to take advantage of these methods. The first stages of geothermal project development require analyzing the data produced by geophysical surveys and integrating heterogeneous datasets. Through AI, regional or small-scale assessment of the geothermal resources assessment can integrate and combine these data and identify suitable locations where proceed with a thoroughly exploration to site geothermal wells and plants that usually require time, invasive inspection, high costs and permissions from legislative authorities. This work presents a worldwide map representing the places with highest suitability for her geothermal wells and plants installation. A Maximum Entropy-based ML model, mediated from ecological niche models, is used to build a spatial probability function based on real-valued vectors of environmental parameters. This function approximates the prior probability that a specific site is suitable for building a geothermal power plant. Our model was trained on environmental parameters in known efficient and operative geothermal power plants. The collected datasets were retrieved from difference repositories and had different spatial distributions and resolutions. Thus, a pre-processing stage was needed for noise and error removal, gap filling and aligning to a 0.5° resolution global grid This phase required algorithms to interpolate and fill data gaps to obtain homogeneous global-scale distributions. To this aim, Inverse Distance Weighting, Kernel Density, and Data-Interpolating Variational Analysis (DIVA) were used. In particular, the collected datasets where: i) distribution of CO2 flux at soil; ii) Digital Elevation Model (DEM); iii) global heat flow; iv) thickness of the sediments; v) average yearly surface air temperature; vi) precipitations; vii) groundwater resources; viii) earthquake depth, magnitude and density distribution; ix) distance from margin plates (e.g., convergent line, transform line, diffuse line, ridge line). The MaxEnt algorithm have been applied to four different sets of the aforementioned datasets: a) to all parameters; b) to a subset obtained by using MaxEnt as filter (i.e., considering only the parameters with a percent contribution higher than 5% of the maximum contribution); c) to a subset of parameters selected by an expert; d) to a combined set coming from MaxEnt filtered and expert-selected parameters. All the four results showed similar patterns but several discrepancies and the suitability score range in the maps were different. The optimal model was the one using MaxEnt-selected features, which correctly predicted the high suitability of 97 geothermal plants over 105 (92%). The produced global suitability map of suitable sites for geothermal plants and wells installation (suitable geothermal sites) is useful to a priori knowledge, to save time and money in the preliminary stage of exploration and when few data are usually available. Moreover, our model can identify the parameter ranges driving a specific site suitability assessment. Thus, it can support communication with citizens whose territories are involved by geothermal probing and power plants installation, who are not usually clearly informed about the scientific reasons driving the selection of their territory and the potential benefits.