Integrating Machine Learning Workflow into Numerical Simulation for Optimizing Oil Recovery in Sand-Shale Sequences and Highly Heterogeneous Reservoir

This paper aims to evaluate the efficiency of various machine learning algorithms integrating with numerical simulations in optimizing oil production for a highly heterogeneous reservoir. An approach leveraging a machine learning workflow for reservoir characterization, history matching, sensitivity analysis, field development and optimization was proposed to accomplish the above goal. A 3D subsurface model representing studied sand-shale sequences was constructed based on geophysical and petrophysical logs, core measurements, and advanced machine learning techniques. After that, a robust sensitivity analysis and history matching process were conducted using a machine learning workflow. The most sensitive control variables were the aquifer properties, permeability heterogeneity in different directions, and water–oil contacts. The history matching results from the constructed geological model showed that the oil rate, water rate, bottom hole pressure, and average reservoir pressure were matched within a 10% deviation from the observed data. Several field development scenarios were generated using the validated model to optimize cumulative oil recovery. Different injection well placement locations, well patterns, and the possibility of converting existing oil-producing wells to water injection wells were investigated. A machine learning-based proxy model was built for the prediction of cumulative oil production and then optimized with hybrid machine learning techniques. The Artificial Neural Network (ANN) algorithm was found to provide higher field cumulative oil production compared with the Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) of 3.5% and 26.5%, respectively. Following the detailed proposed machine learning-based workflow, one can effectively decide on the development strategy and apply the findings from this research to their field.