Generative geomodeling based on flow responses in latent space

This paper presents a new deep-learning-based generative method applicable to history matching without an inverse scheme. Multiple-point geostatistics is used to construct a prior population stochastically. A convolutional variational autoencoder (VAE) with probabilistic latent space is trained as the generative method, and k-means clustering, nondominated sorting, and multilevel geomodel generations are performed based on flow responses. The applicability of the developed workflow was confirmed using a waterflooding problem with multiple wells in fluvial channel reservoirs. The VAE generates new geomodels based on the latent features and builds equiprobable models neighboring the representative models that reflect the observed production performance. The geomodels match the oil production profiles reliably as the steps progress and accurately forecast the water breakthrough time and liquid production trajectories. The density map of plausible geomodels explains reasonably the uncertainty of channel connectivity. The structural similarity index confirms that the generated geomodels become similar to the target reservoir and thus that the developed VAE-based framework creates geomodels that preserve geological realism. This proposed method involves relatively less time-consuming simulations without any inverse or optimization processes; nonetheless, it generates plausible geomodels in dimensionality-reduced latent space. The study methods and findings are thus applicable to scale-variant data integration and uncertainty assessment. •A generative geomodeling framework classified by flow responses was developed.•A convolutional variational autoencoder (VAE) generates plausible geomodels.•The VAE contributes to preserving geological realism.•The workflow is applicable to history matching problems without inverse schemes.