Lawrence Berkeley National Laboratory

The process of going from course-scale seismic reservoir parameters produced from AVA inversion to a fine-scaled reservoir permeability model that fits production data usually results in a permeability and associated seismic parameter model that fits production data but not the original input seismic data. Rarely an iterative process is employed that attempts to find a model that fits both the seismic and production data, but even when successful this is a very expensive. We develop and demonstrate a process that incorporates AVA stochastic inversion with machine-learning to produce fine-scale permeability (and associated seismic parameter) models that fit both the observed seismic AVA and the production data. The process involves training a cGAN on synthetic flow-AVA models to generate a conditional probability function for find-scaled permeability given course-scaled seismic parameters and applying this to the stochastic ensemble of course-scaled AVA inversion models. We show that the resulting MAP permeability model fits production data significantly better than permeability derived from the original AVA models. To further improve production data fit the ensemble of permeability models can be flow-simulated and the closest match to production data chosen to provide the ultimate solution that fits both seismic and production data.