Geospatial Machine Learning Framework for Subsurface Property Mapping: Application to Reservoir Quality Assessment and Carbon Storage

Traditional geostatistical methods face critical limitations in sparse-data environments, including frequent over-smoothing, challenging uncertainty quantification, and limited capability to integrate heterogeneous multi-source datasets for spatial property prediction. Conversely, pure machine learning approaches lack essential spatial coherence required for subsurface applications. This study presents a hybrid framework combining machine learning’s superior data integration capabilities with geostatistical spatial modeling.

The framework supports multiple machine learning algorithms with NGBoost utilized as the primary model due to its superior uncertainty quantification and predictive performance. The methodology effectively integrates diverse geological contexts including depositional environments (GDE), well log summaries, fault networks, paleoburial depth, and structural features, enabling complex pattern recognition beyond traditional geostatistical capabilities. Spatial cross-validation with automated bayesian optimization ensures model reliability, while Gaussian Random Field simulation preserves spatial coherence. Monte Carlo uncertainty propagation enables multi-property integration for reservoir quality assessment.

In a validation case study, the framework was applied to 1,679 wells across the Norwegian and UK Continental Shelves to estimate reservoir properties in potential carbon storage sites. The methodology achieved 35–80% RMSE reductions compared to geostatistical methods, with 84% MAE reduction for net-to-gross prediction and well-calibrated uncertainty estimates (71.2% coverage). These results illustrate its operational readiness for exploration and carbon storage applications.