A global energy operator in the Middle East replaced zone-level manual vug picking with an AI-native computer vision pipeline that quantifies individual vugs — area, circularity, azimuth — at 0.1 m resolution across 200+ m of carbonate borehole image logs, recovering secondary porosity features experts missed and turning qualitative interpretation into structured reservoir geometry.
At a glance
Three metrics frame the shift from manual zone-level picking to automated contour-level quantification.
Depth resolution
Missed vugs recovered
Parameters retuned per new well
The challenge
Carbonate reservoirs store hydrocarbons in fractures, vugs, and inter-crystalline pores — a heterogeneity that matrix-porosity models alone cannot resolve. Secondary porosity — the voids created by dissolution, dolomitization, or karst processes — often controls flow capacity and sweep efficiency, yet traditional manual interpretation from borehole image logs captures only zone-level averages.
The operator's petrophysicists were working with FMI and CMI static logs spanning over 200 m per well in an Oman carbonate field. Manual picking in WellCAD yielded porosity percentages for broad intervals but missed individual vug geometry: size, shape, azimuthal distribution, and vertical clustering. The logs held the signal — dark resistivity anomalies marking fluid-filled voids — but human interpretation was fundamentally limited by pixel-scale resolution, tool variability (FMI depth of investigation ~30 inches vs. CMI ~0.90 inches), and the sheer repetitive effort required to trace thousands of contours.
Without vug-scale quantification, reservoir models treated secondary porosity as a statistical smear rather than discrete geometric objects. The team needed a way to see what experts could not: every vug, its circularity, its orientation, its neighbors — at resolution finer than any manual workflow could sustain.
What we did
We built an AI-native computer vision pipeline that treats borehole image logs as structured image stacks rather than depth plots — segmenting raw FMI and CMI static logs into 1 m processing zones, each decomposed into 0.1 m reporting intervals, and applying a five-stage detection cascade tuned to carbonate vug morphology.
The pipeline starts with top-k mode subtraction, removing the k=5 most frequent pixel intensity values to suppress banding artifacts and tool noise. Then Gaussian-modulated adaptive thresholding isolates candidate dark anomalies — the resistivity signature of fluid-filled vugs — using a locally adaptive threshold rather than a global cutoff, preserving vugs in both high- and low-resistivity zones.
Five-stage vug detection cascade
Top-k mode subtraction
k=5 to suppress banding
Gaussian-adaptive thresholding
Locally adaptive, zone-specific
Contour extraction
Circularity filter 0.3–1.0
Dual-layer filtering
Laplacian edge + mean deviation
Vug catalog output
Area, circularity, azimuth, resistivity
Contour extraction follows, filtered by circularity (0.3–1.0) to reject linear fractures and wellbore-parallel artifacts. Each contour is then subjected to dual-layer filtering: Laplacian edge strength to confirm sharp boundaries, and mean resistivity deviation to reject noise patches. What survives is a vug: a discrete geometric object with measured area (1–12 cm² in this field), circularity (0.28–0.85), and azimuthal position across four quadrants.
We deployed the pipeline across three wells — two vertical, one horizontal — with varying tool types (FMI in Wells A and B, CMI in Well C) and resistivity ranges. The processing stack required only 2–3 parameter adjustments per new well (ΔL for brightness normalization, Δμ for mean deviation threshold, and adaptive block size), preserving the core detection logic across tool and orientation shifts.
The output is not a zone-level percentage but a vug catalog: depth, area, circularity, azimuth, resistivity profile — every 10 cm, over 200+ m. We cross-validated against manual WellCAD picks and identified at least four discrete depth intervals across Wells A and B where the pipeline recovered vugs the expert interpreter had missed — vugs that met every geometric and resistivity criterion but fell below the threshold of manual attention.
The outcome
The pipeline processed 200+ m of log per well and delivered vug-level quantification at 0.1 m resolution — a 2,000-fold refinement over the zone-level (tens of meters) granularity of manual picking. Individual vugs ranged from 1 to 12 cm² in area, with circularity spanning 0.28 (elongated, dissolution-aligned) to 0.85 (near-circular, isotropic voids).
Cross-validation against manual WellCAD interpretation revealed that the pipeline recovered vugs in at least four discrete depth intervals across Wells A and B that the expert had missed — not because the signal was absent, but because pixel-scale manual tracing cannot sustain attention across thousands of candidate anomalies. The automated contour filter applied the same geometric and resistivity criteria uniformly, without fatigue or subjective thresholding drift.
Parameter retuning for new wells was minimal: only 2–3 of the 10+ tunable parameters (ΔL for brightness normalization, Δμ for mean deviation threshold, adaptive block size) required adjustment when moving from FMI to CMI logs or from vertical to horizontal orientations. The core detection logic — top-k subtraction, Gaussian-adaptive thresholding, circularity filtering — generalized across tool types and well geometries without retraining or architecture changes.
What changed
- Depth resolution improved from zone-level (tens of meters) to 0.1 m intervals — 2,000× finer granularity
- Four discrete depth intervals showed vugs recovered by automation that manual picking missed
- Minimal retuning (2–3 parameters) enabled deployment across FMI, CMI, vertical, and horizontal wells
What this unlocked
Vug-level geometry — not zone-level porosity percentages — is what reservoir simulators need to honor flow heterogeneity. The catalog of per-vug area, circularity, azimuth, and vertical clustering feeds directly into discrete fracture network (DFN) models, allowing engineers to condition flow simulations on measured void architecture rather than statistical averages.
Azimuthal distribution across four quadrants revealed systematic vug alignment in Wells A and B — a dissolution fabric invisible in zone-averaged picks but critical for understanding preferential flow pathways and optimizing completion design. In Well C (horizontal, CMI), the pipeline quantified vug density variation along the lateral, informing stage spacing and perforation placement.
The workflow's generalization across tool types and well orientations means the operator can now process legacy FMI archives and live CMI logs with the same detection stack, building a field-scale vug catalog without per-well model retraining. Secondary porosity, once a qualitative nuisance term in petrophysical models, becomes a structured data asset — searchable by geometry, depth, and azimuth.
Lessons and next steps
The richest signal in borehole image logs is not binary presence-or-absence but geometric detail: vug shape, orientation, clustering. Manual interpretation, constrained by pixel-scale attention and subjective thresholding, cannot capture that detail at scale. Computer vision, tuned to carbonate morphology and backed by adaptive filtering, can — and does so uniformly across hundreds of meters and multiple tool modalities.
The pipeline's minimal retuning burden (2–3 parameters per new well) suggests that the core detection logic has captured something generalizable about vug resistivity signatures, not overfitted to a single tool or field. The next deployment will test this hypothesis on a different carbonate play — Jurassic Arab formations in Saudi Arabia — where vug size distribution and circularity profiles differ from the Oman dataset.
Automated contour-level quantification doesn't just speed up what experts already do — it sees what they cannot.
Before
Zone-level
After
0.1 m intervals
