Harnessing Generative AI to Enhance Reservoir Property Modeling with Uncertainty Quantification

This study presents a novel workflow for reservoir property modeling using generative artificial intelligence (AI), integrating Generative Adversarial Networks (GANs) and diffusion models to predict porosity and permeability distributions with improved resolution and uncertainty quantification. Traditional geostatistical methods often face limitations in capturing heterogeneity and thin‑bedded features, especially in data‑sparse environments. The proposed approach leverages multi‑source datasets, including well logs, core measurements, and seismic attributes, combined with physics‑informed constraints derived from seismic inversion. A conditional GAN framework generates high‑resolution three‑dimensional property cubes, while diffusion models provide multiple equiprobable realizations to account for geological uncertainty.

Results from synthetic and field‑scale case studies demonstrate significant improvements compared to conventional methods, including enhanced thin‑bed continuity, facies distribution alignment with seismic reflectors, and up to a 20% reduction in mean squared error between predicted and core‑measured porosity. This workflow provides interpretable and uncertainty‑aware reservoir models, supporting more informed decision‑making in exploration and production. The integration of generative AI with geophysical and petrophysical data highlights its potential as a transformative tool for reservoir characterization, with future scope for coupling with digital twins for real‑time adaptive reservoir management.