InterPore2026

Stochastic Upscaling of Hydraulic Properties in Natural Shear Fractures

20 May 2026, 09:20

15m

Oral Presentation (MS19) Uncertainty-Aware Decision Support in Porous Media Applications MS19

Speaker

Dr Sarah PEREZ (Heriot-Watt University)

Description

Accurate prediction of $CO_2$ storage performance in fractured geological formations depends critically on how uncertainty is transferred from the scale of individual fractures to the reservoir grid scale. Natural fracture networks exhibit complex aperture variability, roughness-controlled flow, and spatially correlated heterogeneity, yet conventional cubic-law representations often fail to capture how these features influence large-scale fluid flow and transport [1,2]. This work develops an uncertainty-upscaling workflow that quantifies how geometric uncertainties propagate to hydraulic response in complex fracture systems.

At the local scale, uncertainties in the fracture conductivity are characterised through Bayesian correction of simplified flow laws, yielding posterior permeability distribution instead of a single-value estimate [3]. This step mitigates model misspecification and produces uncertainty-aware training data that reflect the variability observed in real fractures. These posterior fields then support a purely data-driven upscaling strategy capable of bridging multiple orders of magnitude in scale.

A U-Net surrogate is trained on paired fracture-image and hydraulic-response datasets to learn a probabilistic mapping from geometry to permeability. Once trained, the model generates distributions of hydraulic properties directly from aperture images, allowing uncertainty to be efficiently propagated to larger resolutions and to more complex fracture systems. The resulting ensembles preserve key structural features such as channelisation, contact zones, and preferential pathways, while retaining fine-scale uncertainty that deterministic upscaling systematically discards.

We demonstrate the developed workflow on natural sheared fractures extracted from a regional caprock formation within a natural $CO_2$ reservoir in Utah [4]. By combining physics-based correction with data-driven upscaling, probabilistic flow predictions are produced at negligible cost relative to direct Monte-Carlo exploration of the fracture geometry, rendering uncertainty quantification tractable for high-resolution fracture systems. Overall, the workflow provides a scalable surrogate for uncertainty-aware predictions at larger scales by converting imperfect geometric observations into actionable hydraulic responses relevant for leakage-risk assessment.