Ensemble history-matching workflow using interpretable SPADE-GAN geomodel

Ensemble history matching adjusts multiple geomodels used for reservoir simulation, conditioning them to historical data. It reduces and quantifies the uncertainty in the unknown model parameters to increase the models’ reliability for decision support. In this study, we adapt the latest generation of generative artificial intelligence algorithms, SPADE-GANs. In geosciences, Generative Adversarial Networks (GANs) learn to simulate complex geological patterns. SPADE (SPatially Adaptive DEnormalisation) layers in the GAN generator learn conditioning to the coarse geological structure provided as coarse-scale maps, enabling explainable output, stable training, and higher variability of resulting outputs. Our statistical method, an iterative ensemble smoother, assimilates data into an ensemble of these maps, interpreted as the channel proportions. This Bayesian data assimilation conditions the ensemble of GAN-geomodels to a combination of well data and flow data, thus extending the usability of pre-trained SPADE-GANs in subsurface applications. Our numerical experiments convincingly demonstrate the method’s capacity to replicate previously unseen geological configurations beyond GAN’s training data. This proficiency is particularly valuable in data-scarce scenarios typical for renewable geo-energy, where the GAN captures realistic geology, but its output geomodels must be adjusted to match observed data. Furthermore, our fully open-source developments lay the foundation for future multi-scale enhancements of history matching workflows.