Machine Learning-Enhanced Play Fairway Analysis for Uncertainty Characterization and Decision Support in Geothermal Exploration

Abstract

Geothermal exploration has traditionally relied on geological, geochemical, or geophysical surveys for evidence of adequate enthalpy, fluids, and permeability in the subsurface prior to drilling. The recent adoption of play fairway analysis (PFA), a method used in oil and gas exploration, has progressed to include machine learning (ML) for predicting geothermal drill site favorability. This study introduces a novel approach that extends ML PFA predictions with uncertainty characterization. Four ML algorithms—logistic regression, a decision tree, a gradient-boosted forest, and a neural network—are used to evaluate the subsurface enthalpy resource potential for conventional or EGS prospecting. Normalized Shannon entropy is calculated to assess three spatially variable sources of uncertainty in the analysis: model representation, model parameterization, and feature interpolation. When applied to southwest New Mexico, this approach reveals consistent enthalpy trends embedded in a high-dimensional feature set and detected by multiple algorithms. The uncertainty analysis highlights spatial regions where ML models disagree, highly parameterized models are poorly constrained, and predictions show sensitivity to errors in important features. Rapid insights from this analysis enable exploration teams to optimize allocation decisions of limited financial and human resources during the early stages of a geothermal exploration campaign.