Orientation-Guided Gaussian Processes for Probabilistic Mapping of Gas Reservoir Effective Thickness

Probabilistic maps of gas reservoir effective thickness are increasingly used in development decisions, but production workflows rarely show how structural guidance carries into predictive uncertainty and threshold-based interpretation. We present a field-scale workflow with two linked outputs: a deployed anisotropic KNN-IDW contour surface derived from interpreted gully polygons, and an anisotropic Matérn-3/2 Gaussian process (GP). The GP audits the orientation prior by REML fitting and yields exceedance probabilities alongside the deterministic contour product. The orientation prior is constructed from polygon geometry only, and its provenance—an independently interpreted regional sedimentary-facies basemap—is external to the thickness attribute being mapped, so the structure is not reverse-engineered from the reported values. In a de-identified benchmark of 414 unique wells with strongly nugget-dominated thickness (74–81% across fitted variants), the log-space REML azimuth lies 5.3° from the geometric gully axis. At this nugget level, freely estimated azimuths become protocol-unstable, so the workflow locks orientation to the prior. The isotropic null is rejected by both a boundary calibration and a 1000-replicate parametric bootstrap. Audited anisotropy improves GP/BLUP prediction—stronger under pointwise-LOO than under spatial-block audit—although the field contrasts remain within bootstrap uncertainty. Under spatial-block validation, the gains are much smaller, and imposing the same constraint on distance-based IDW brings little point-accuracy benefit. The exceedance maps classify 96.9% of the valid area as decision-uncertain, with the decisive low-probability region (1084 cells, 3.09%) falling entirely below the deterministic 2 m contour and zero cells in the flipped direction, so the probabilistic layer is best read as a tool for risk and decision triage—exposing where the deterministic contour is comparatively unsupported—rather than as a replacement for the deterministic contour with one definitive probability boundary. Overall, the results support an auditable, uncertainty-aware workflow while keeping the limits in spatial generalization and cross-reservoir transfer explicit.