Parametrization of stochastic inputs using generative adversarial networks with application in geology

摘要: We investigate artificial neural networks as a parametrization tool for stochastic inputs in numerical simulations. address from the point of view emulating data generating process, instead explicitly constructing parametric form to preserve predefined statistics data. This is done by training network generate samples distribution using recent deep learning technique called generative adversarial networks. By relevant are replicated. The method assessed subsurface flow problems, where effective underground properties such permeability important due high dimensionality and presence spatial correlations. experiment with realizations binary channelized perform uncertainty quantification parameter estimation. Results show that very preserving visual realism well order responses, while achieving reduction two orders magnitude.