Bridging Reservoir- and Pore-Scale Modeling with Pix2Pix cGANs

Summary Generative artificial intelligence offers powerful new tools for subsurface characterization, complementing traditional geostatistical methods that often fail to capture complex heterogeneity in carbonate systems. This study applies a Pix2Pix conditional generative adversarial network (cGAN) to perform image-to-image (property-to-property) translation across reservoir and pore scales. At the reservoir scale, a 7.6 million-cell dataset from the Groningen Rotliegend field was used to train a U-Net generator with a PatchGAN discriminator over 19 000 iterations. The model achieved high predictive accuracy for facies (PA 0.88; FWIoU 0.85) and water saturation (PA 0.96; FWIoU 0.95), and strong cross-property translation performance (facies → Sw: PA 0.98; FWIoU 0.97). Variogram-based validation confirmed geological realism. At the pore scale, the same framework reconstructed fine-scale heterogeneity, connectivity, and multiphase-flow pathways from coarse pore structures, reproducing ground-truth simulations. These findings demonstrate that cGAN-based translation can generalize from siliciclastic to carbonate systems, enabling scalable, data-driven modeling of heterogeneity critical to reservoir management and energy-transition applications such as geothermal, CO2 storage, and water-resource studies. The work contributes to the “Machine Learning and Digital Transformation in Carbonates” theme by showcasing a unified AI approach bridging reservoir and pore-scale characterization.