Machine Learning Optimization of Geothermal Potential Prediction with Key Control Quantification: A Case Study from Alberta's Basal Cambrian Sandstone Unit

Abstract

Accurately predicting geothermal potential in sedimentary basins is critical for de-risking exploration. This study develops a robust machine learning (ML) framework that prioritizes predictive integrity through rigorous model benchmarking and interpretability analysis. Using the Lower Cambrian sandstone in the Alberta Basin as a case study, we demonstrate a workflow that not only predicts geothermal potential but also quantifies prediction uncertainty and identifies key controlling factors. A comprehensive dataset from 301 wells, integrating key geological (Reservoir Temperature, Effective Geothermal Reservoir Thickness, Porosity) and engineering (Extraction Depth, Utilization Factor, location) parameters, is used to train and test models predicting single-well Total Heat Capacity (MW·h/m2). A rigorous benchmark test protocol, including train-validation-test splits and an independent blind test set, was employed to ensure fair comparison and assess generalization. The results show that the XGBoost algorithm achieved superior performance on the hold-out test set, with a coefficient of determination (R2) of 0.879 and a root mean square error (RMSE) of 0.13. Its residuals approximated a normal distribution, indicating excellent model robustness. In contrast, simpler models like DT and SVR demonstrated significantly lower predictive capability for this complex, non-linear problem. Beyond prediction, we employ Shapley Additive exPlanations (SHAP) analysis to interpret the optimal XGBoost model and quantitatively deconstruct the influence of input features. This interpretability approach identifies Extraction Effective Geothermal Reservoir Thickness, location, and Reservoir Temperature as the three dominant controlling factors, collectively accounting for over 86.7% of the model's predictive output. SHAP dependency plots further elucidate the specific, often non-linear, relationships between these key factors and geothermal potential. The finalized model validated its practical reliability on a completely independent blind dataset, achieving a mean absolute percentage error (MAPE) of less than 0.08. This work contributes a validated, end-to-end ML workflow that transitions from systematic algorithm selection to actionable geological insight. The framework offers a practical solution for the rapid assessment and sweet-spot identification in sedimentary geothermal plays, thereby aiding in resource characterization and investment decision-making.