Machine Learning Enables Data-Driven Predictions of CO2 EOR Numerical Studies

Abstract

Listen

Translate article

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 218441, “Data-Driven Predictions of CO2 EOR Numerical Studies Using Machine Learning in an Open-Source Framework,” by Jorge R. Liguizaca, SPE, University of Bergen and Escuela Superior Politécnica del Litoral, and David Landa-Marbán and Sarah E. Gasda, NORCE, et al. The paper has not been peer reviewed. _ An open-source framework is presented for the development and evaluation of machine learning- (ML) assisted data-driven models of CO2 enhanced oil recovery (EOR) processes to predict oil production and CO2 retention. The main objective of the authors was to increase the speed, robustness, and accuracy of predicting oil recovery and CO2 retention using a complete open-source approach combining Python programming, reservoir simulation, and ML techniques. Overview of the Open-Source Framework The evaluation of the predictive models was performed using two CO2 water-alternating-gas (WAG) simulation cases, which were proposed using the SPE Comparative Solution Project (CSP) 5 simulation model as a reference. First, a reservoir simulation deck template and a configuration file, including variable inputs, were generated to create any number of simulation jobs for each case. The input-value range for the simulations was set in a Python script. Then, the models were generated and run using the Python package pyopmnearwell and the open-source reservoir simulator OPM Flow to determine oil production and CO2 retention. Moreover, a Python script was developed for training, testing, validating, tuning, and deploying predictive models using the inputs and output data from the reservoir simulation and scikit-learn, an open-source framework for ML, for comparing the results of the simulations with the predictive models. Base-Model Creation. The first step for base-model creation was the development of the reservoir simulation deck template. This model took as a reference the SPE CSP 5 reservoir model, which consists of a 3D domain grid with an injector well and a producer well. The expanded model grid is 7×7×3, containing 147 grid cells. The gridblocks are squares in x and y coordinates, measuring 500 ft by 500 ft, with each layer’s thickness ranging from 20, 30, and 50 ft from top to bottom. The horizontal permeability for each layer is 500, 50, and 200 md, respectively, whereas the vertical permeabilities are 50, 50, and 25 md, respectively. The geological model has a homogeneous porosity of 30%. The injector well is situated at the bottom left corner and can inject CO2 or water. The producer is located at the top right corner, with a maximum production objective of 12,000 STB/D. This work was developed to evaluate the predictive models’ performance, comparing them with reservoir simulation to assess the effects of a set of inputs on oil production and CO2 retention outputs. The selected inputs were CO2 injection rates and the porosity and permeability in the top layer. The top-layer-grid properties were selected as a variable parameter. Because the CO2 injection rate affects the sweep efficiency and pressure maintenance, assessing its effect on CO2 storage and oil production was essential.