|
|
| |
Explainable 3D EGS Targeting Using Self-Organizing Maps and Integrated Geoscientific Proxies
Zeynep ANKUT AYDAR, Sebnem DUZGUN
[Colorado School Of Mines, USA]
Enhanced Geothermal System (EGS) target identification requires integration of multimodal geoscientific data to capture thermal state, rock properties, and structural controls across three-dimensional (3D) subsurface volumes. Understanding the non-linear relationships between EGS favorability proxies is essential for targeted data collection and informed prospectivity analysis, yet traditional approaches lack robust mechanisms to reveal these critical interdependencies. This study presents a framework for EGS characterization that combines Self-Organizing Maps (SOM) with depth-resolved statistical validation to distinguish EGS from naturally conventional hydrothermal systems. We developed integrated 3D subsurface models at Utah Frontier Observatory for Research in Geothermal Energy (Utah FORGE) and Roosevelt Hot Springs (RHS) comprising 4.5 million blocks at 100 m resolution, incorporating 11 attributes: P-wave velocity (Vp), S-wave velocity (Vs), Vp/Vs ratio, density inferred from 3D magnetic gravity inversion, temperature, fault density, and five lithology indicators. A 3×3 SOM was trained using a representative subset of the 4.5M valid blocks and then applied to the full block volume to obtain 3D cluster assignments. Component plane analysis revealed systematic correlations: high-temperature domains consistently co-locate with elevated Vp/Vs ratios and intermediate densities, confirming coupling between thermal anomalies, fluid presence, and enhanced permeability. Within the RHS footprint, Cluster C7 represents the dominant high-temperature regime and shows negligible within-cluster sampling bias (Δmean ≈ 0). This supports that the footprint-labeled blocks are representative of the underlying SOM prototype. Depth-resolved statistical validation using dual parametric (Welch's t-test, Hedges’ g) and non-parametric (Mann–Whitney U, Cliff's δ) frameworks quantified system-level contrasts across 500 m depth bins. RHS exhibits very large thermal effect sizes at shallow depths (g greater than 2, |δ| greater than 0.7) reflecting strong convective upflow, while FORGE shows comparable or elevated temperatures below 3000 m depth consistent with deep EGS targeting. This work establishes a reproducible, objective framework for 3D EGS prospectivity assessment that eliminates subjective weighting, delivers quantitative explainability of proxy interdependencies, and provides spatially explicit volumetric delineation of prospect quality. Unlike black-box approaches, the SOM architecture preserves interpretability while handling high-dimensional data, enabling geoscientists to interrogate proxy relationships, validate assignments against domain knowledge, and extract transferable insights. Explainability is provided via SOM component planes (proxy co-variation), cluster prototypes, and depth-binned effect sizes that quantify where and why system contrasts occur. The integrated methodology advances EGS exploration by demonstrating that combined unsupervised learning and rigorous statistical validation can identify subsurface targets in structurally complex basement environments while maintaining full transparency of decision criteria.
Topic: Enhanced Geothermal Systems