A stochastic learning-from-data approach to the history-matching problem

History matching is the process whereby the values of uncertain attributes of a reservoir model are changed with the purpose of finding models that match existing reservoir production data. As an inverse and ill-posed problem in engineering, it admits multiple solutions and plays a key role in reservoir management tasks: reservoir models support important and strategic field development decisions and, the more calibrated the models, the higher the confidence on their forecast for the actual reservoir’s performance. In this work, we introduce a stochastic learning-from-data approach to the history-matching problem. With a data-driven nature, the proposed algorithm has dedicated components to handle petrophysical and global uncertain attributes, and generates new solutions using the patterns of attributes present in solutions that are judiciously selected among a set of solutions for each well and variable involved in the history-matching process. We apply our approach to the UNISIM-I-H benchmark, a challenging synthetic case based on the Namorado Field, Campos Basin, Brazil. The results indicate the potential of our learning proposal towards generating multiple solutions that not only match the history data but, most importantly, offer acceptable performance while forecasting field production. Compared with history-matching methodologies previously applied to the same benchmark, our approach produces competitive results in terms of matching quality and forecast capacity, using substantially fewer simulations.