The European Conference on the Mathematics of Oil Recovery (ECMOR) is one of the most important events that gathers worldwide experts from Oil & Gas Technology teams and academia researchers to exchange ideas on reservoir characterization, simulation and optimization.
Our CEO, Hector Klie, had the opportunity to co-chair sessions on physical modeling and discretization and present the work Data-Driven Modeling of Fractured Shale Reservoirs in the session “Machine Learning and Proxy Methods”. In this work, jointly developed with Horacio Florez, Texas A&M, Dr. Klie introduces a novel data-driven methodology based on the Dynamic Mode Decomposition (DMD) approach to infer the dynamics entailed by coupled flow/geomechanics on fractured media.
Results were based on full-physics simulations and establish accurate pressure field inference from an unseen combination of fractures, fracture length and permeability values. The computational experiments revealed that the proposed predictive approach is 20 times faster than a deep learning convolutional network achieving similar accuracy. Moreover, each DMD prediction is 500 times faster than a high-fidelity simulation.
Regarding production profiles, it was shown that the Long Short-Term Memory (LSTM) approach was very accurate to infer production profiles. The computational cost implied by LSTM was marginal compared to the pressure field inference. Despite the preliminary character of the results for coupled single-phase flow and geomechanics, there are indications for achieving important computational savings for performing modeling and optimization on unconventional assets.