Application of machine learning techniques for modeling of relative permeability in engineered water injection in carbonate reservoirs.

Abstract

Numerical modeling of advanced production methods is always a challenge to be developed and applied in reservoir simulation. Some approaches, such as the use of laboratory experiments, arise to make this modeling feasible. However, this limits the speed of the solution to obtaining laboratory data and impairs its reproducibility. With the increasing use of Machine Learning (ML) tools to solve complex non-linear problems, we conducted these studies to train these ML tools and couple them to commercial simulation software. The training was based on parameters relevant to Engineered Water Injection (EWI). This advanced injection method seeks to use salinity control in the injection water to promote iterations between its ions and the rock minerals, to facilitate its flow into the porous medium. Thus, we structured a dataset containing salinity, mineralogy, and relative permeability data for the data-driven ML tool to learn the behavior of this data.Thus, this approach achieves accurate predictions, which were used as input data during injection modeling and simulation, validating its results by comparing with production simulation by conventional geochemical modeling. Finally, we performed optimizations with waterflooding injection and EWI, coupling the optimization with the advanced method of the ML pipeline. Thus, we test the efficiency of the ML approach with recursive simulations and compare the efficiency between the injection methods. For this, we apply these optimizations to the UNISIM-II benchmark, a reservoir model with characteristics based on Brazilian Pre-Salt fields. The objective function was Net Present Value maximization, which for the tests performed, EWI presented higher profit, even with a cost margin up to 300% higher than the cost of waterflooding.