Resolution-aware multiscale 3D CNNs for permeability estimation in carbonate rocks: an architecture-agnostic transfer-learning framework

Abstract

Permeability is a key parameter controlling groundwater flow, contaminant transport, and subsurface storage, yet its estimation from imaging data is hindered by the trade-off between micro-CT resolution and field of view (FOV). High-resolution scans capture critical pore-scale features but only over small volumes, whereas coarse images cover representative domains but lack sufficient detail for direct numerical simulation. Building on our previous proof-of-concept resolution-aware 3D convolutional neural network (3D CNN) surrogate for tight carbonate rock (Nabipour et al. (2024)), we develop and generalize this framework by introducing a compact Multiscale Resolution-Aware Network (MRAN) and systematically benchmarking it against three established 3D CNN architectures: ResNet50, ResNeXt50, and DenseNet201, using multiresolution micro-CT images of tight carbonate rock and lattice Boltzmann (LBM)-derived permeabilities. All networks are trained with a fine, intermediate and coarse (FR-IR-CR) hierarchical transfer-learning scheme on two datasets containing 1416 and 3600 distance transform and binary multiscale subvolumes, enabling a controlled assessment of architecture choice, model complexity, and data scaling. Results show that hierarchical transfer learning is architecture-agnostic and consistently improves coarse-resolution performance, with all models achieving R² ≥ 0.94 at a 3 µm voxel size. MRAN contains fewer than ∼1.2 million trainable parameters, ≈ 2 % of ResNet50’s, and achieves R² > 0.97 on coarse-resolution validation data, outperforming the deeper backbones by 30–45 % .Validation against independent mercury-intrusion porosimetry measurements on unseen coarse-resolution volumes yields relative permeability errors as low as 7.5%, and MRAN best reproduces the full permeability distribution, including both high- and low-permeability tails of the distribution, whereas the baselines exceed +12 %. Once trained, inference on 600³-voxel images takes only 10 min, providing orders-of-magnitude speedups over direct simulation. The proposed architecture-agnostic and resolution-aware workflow provides a practical route to scalable permeability estimation in heterogeneous carbonate reservoirs and is conceptually applicable to other porous media and transport properties.