Approaching geoscientific inverse problems with vector-to-image domain transfer networks

• We propose a deep neural network to predict 2D subsurface property fields from 1D measurement data (e.g., time series) thereby offering new way perform inversion. The method is illustrated using inversion of synthetic (1) first-arrival GPR travel times and (2) series transient hydraulic heads multi-pumping experiment. generalizes by producing more appropriate models when fed with than those found in the training set. Uncertainty assessed ensemble. present vec2pix , designed categorical or continuous one-dimensional series), an alternative approach solve inverse problems. performance investigated through two types problems: (a) crosshole ground penetrating radar (GPR) tomography experiment being used infer velocity field, multi-well pumping within unconfined aquifer retrieve conductivity field. For each type problem, both multi-Gaussian binary channelized domain long-range connectivity are considered. Using set 20,000 examples (implying as many forward model evaluations), recover that much closer agreement true closest forward-simulated space. Further testing smaller sample sizes shows only moderate reduction 5000 only. Although these runs can be performed offline parallel, associated computational expense may still prohibitive for very demanding solvers. In addition, re-training required configuration. Even if recovered visually close ones, misfits their responses generally larger noise level contaminate data. solution partially ensembles, which trained repeatedly random initialization. Overall, this study advances understanding how use learning indirect More work needed evaluate suitability our proposed real data, errors uncertainty geologic scenario will bring additional complexities.