InterPore2026

Connectivity-aware pore segmentation in carbonate SEM images using an attention U-Net with physics-aware refinement.

19 May 2026, 09:50

1h 30m

Poster Presentation (MS15) Machine Learning in Porous Media Poster

Speaker

Wurood Alwan (University of Leeds)

Description

Carbonate reservoir performance depends not only on total porosity, but on how pore-space connectivity controls transport. In backscatter SEM (BSE–SEM) images, connected pathways and isolated intragranular pores can have similar greyscale appearance, yet they imply very different behaviour: connected pores support flow, whereas isolated pores mainly contribute to storage and trapping. Standard “pore vs. matrix” segmentations, therefore, risk biasing permeability proxies and connectivity descriptors when all pores are treated equivalently.

This ongoing work produces connectivity-aware pore maps from 2D carbonate BSE–SEM by distinguishing three phases: isolated intragranular pores, connected pore pathways, and mineral matrix. The workflow is demonstrated on four large SEM mosaics (29,056 × 22,952 px; 0.195 µm/px) partitioned into 2048 × 2048 px tiles (100 labelled tiles), with evaluation on a strict held-out test set of 20 tiles. Connectivity labels are derived from rapid grain-boundary (yellow ring) annotations that separate pores inside grains from pores outside grains while preserving thin (1–2 px) throats and filamentary links.

On the held-out test set, the approach reproduces the mineral matrix with high overlap (IoU ≈ 0.92) and delineates the connected pore network with moderate-to-strong overlap (IoU ≈ 0.48). Isolated pores are extremely rare (≈0.039% of test pixels) and remain the most challenging class, but their detection improves after a light refinement step (IoU 0.035 → 0.069; recall 0.18 → 0.30) while the connected-pore and matrix classes change only marginally.

These connectivity-aware masks enable direct quantification of connected versus isolated porosity fractions and provide inputs compatible with downstream digital-rock connectivity analyses (such as topology- and percolation-inspired descriptors). This is particularly relevant to subsurface applications where the balance between mobile and trapped porosity controls long-term performance, including CO2 storage and radioactive-waste disposal.