Mining Legacy Geothermal Databases with Unsupervised Learning for Dual-Resource Exploration

Abstract

Legacy geothermal datasets provide a rich yet underutilized resource for advancing data-driven subsurface exploration. We applied unsupervised machine learning to a digitized U.S. Geological Survey database containing over 1,800 water and 300 gas chemistry samples (1930–2006) from geothermal and hot spring systems across the western United States. The objectives were twofold: (i) to delineate geothermal provinces and fluid sources, and (ii) to identify sites with geochemical conditions conducive to geological hydrogen generation through serpentinization of ultramafic rocks. A multi-stage analysis was performed using k-means, DBSCAN, and Gaussian Mixture Models on standardized chemical, isotopic, and temperature data. Geochemical fingerprinting revealed coherent clusters representing magmatic–volcanic, sedimentary, and deep-circulation systems, providing analogs to productive geothermal fields and flagging unexplored zones. Cluster tagging based on high pH–Mg waters, reducing conditions, and mantle gas indicators (e.g., elevated ³He/⁴He) identified candidate ultramafic settings favorable for H₂ formation and preservation. Hydrochemical–thermal zonation further distinguished high- versus low-enthalpy systems and reconstructed fluid mixing trends within serpentinization-relevant temperature and redox windows. Temporal analyses across seven decades corrected historical biases and improved confidence in H₂ prospectivity screening. This framework illustrates how legacy datasets, when integrated with unsupervised learning, can delineate geothermal provinces, track geochemical evolution, and identify ultramafic terrains for exploration and engineering of hydrogen generation systems. The results highlight the untapped potential of historical data to accelerate sustainable subsurface energy discovery.