How selecting representative training sites and features can improve machine learning models of hydrothermal resource favorability

Fundamental challenges for improving supervised machine learning (ML) models for predicting geothermal resources include: 1) available training data being mostly positive (e.g., hot springs or operating power plants), so negative training sites must be selected, and in particular, negative training sites must represent the full range of conditions where no resources exist; and 2) input features potentially containing signals of several geologic processes, some of which may not be directly informative for modeling the availability of geothermal resources. We consider the artificial neural network (ANN) developed by the Nevada Machine Learning project (NVML), propose new selection strategies for negatives, and use signal separation to enhance the information content of input features. Using these strategies, we develop ML models using arguably simpler algorithmic approaches (Logistic Regression and eXtreme Gradient Boosting [i.e., XGBoost]) and compare the new results with the original ANN results. In particular, we evaluate the influence of selection strategy for negative training sites by comparing two training strategies: 1) using the original negative sites defined by NVML (NVML negatives); and 2) randomly selecting negatives from outside ellipsoids defined by NVML as having favorable structural settings, under the assumption that areas far from favorable settings are likely negative. Next, we explore the added value of splitting the topography input feature into two features to separate signals into regional elevation trend and local relative elevation, under the assumption that regional trends are correlated to regional geology (e.g., crustal thickness) and that local relative elevation (e.g., basin and range topography) is correlated to local geology (e.g., faulting). All new models using logistic regression and XGBoost were run 60 times each using Monte Carlo cross-validation to evaluate predictive uncertainty. Models using the randomly selected negatives outperform models using the negatives selected for the NVML ANN; models using XGBoost outperform models using logistic regression; and models using separated elevation features outperform models using the original elevation data. The results showing improved performance with random sampling of negatives demonstrates the value of sampling the entire input feature space when fitting a model (i.e., providing the ML workflow with all types of negatives and positives). The results showing improved performance with separated topography demonstrates that ML models cannot always find important patterns when they are obscured by larger trends, so feature engineering that enhances these patterns can be beneficial.