Machine-Learning Application for Fast 3-Phase Black Oil Reservoir Simulation

Reservoir simulation is essential for field development and well planning through the estimation of pressure, saturation distribution, and production rate decline prediction. One of the primary objectives of the simulation is to predict the hydrocarbon recovery from a reservoir in a given period of time under study. When the field-scale reservoir model is highly heterogeneous and complex, then the commercial reservoir simulators employ a large number of grid blocks to model the reservoir, and this necessitates the solving of a large number of finite-difference discretized mass balance equations. Solving these equations requires a large amount of computational time and detailed information about the reservoir properties like porosity and permeability. Also, when the fluid properties are highly sensitive to pressure, then a large number of iterations are performed at each time step to predict the fluid properties, which as a result, makes the traditional reservoir simulation process computationally expensive. In this article, a novel methodology is proposed in which a machine learning (ML) model designed using a random forest is used to reduce the number of iterations performed at each time step, which expedites the simulation process. This is done by predicting reservoir fluid properties like formation volume factor, density, and viscosity using ML at the first iteration, which provides a close initial estimate of the parameters needed to solve the finite-difference equations. This first step reduces the number of iterations required to reach convergence and solve for pressure and saturation without changing the conventional finite-difference equations. The applicability of the proposed methodology has been verified on a 3-phase black oil reservoir model. The study’s novelty is in the development of a novel ML-based 3-phase black oil reservoir simulator that can speed up the reservoir simulation process while conserving the standard fluid flow physics involved, making the model computationally more efficient.