THE USE OF DEEP LEARNING WITH PROGRESSIVE GROWING GENERATIVE ADVERSARIAL NETWORKS FOR CONNECTIVITY MODELING IN COMPLEX AQUIFERS

Accurate characterization of connectivity plays a key role for water resources management and aquifer remediation. Unlike traditional multiple-point geo-statistics, this study applies deep learning to depict the preferential paths (i.e., connectivity) based on the training data. Specifically, we focus on progressive growing generative adversarial networks (PGGANs) to condition on measured data (i.e., hard data). Given a latent variable and an array that provides hard data, the generators of the conditioned PGGANs are tasked to produce geologically realistic images of channel aquifers that match with field observations. To better understand the conditioning mechanism, the conditioning behavior of these networks were measured using the conditioning ratio, which is a metric defined to measure the magnitude of the influence of the conditioning input. The conditioning ratio was measured across multiple layers within the generator during training, as well as with various modifications to the network architecture. The results reveal two distinct conditioning behaviors that are based on the number of conditioning arrays included into the generator. Results also show that lowering the starting resolution for the generator can slow down the learning process. Overall, the examples demonstrate that PGGANs can be successfully used for characterization of connectivity in aquifer modeling.