Machine Learning-Based Waterflood Optimization System for Improved Oil Recovery

Abstract

Although waterflooding remains the primary method of secondary oil recovery, its efficiency is often limited by reservoir heterogeneity, inefficient injection allocation, non-uniform sweep efficiency, and uncertainty in injector-producer connectivity. Traditional full-physics reservoir simulation models are computationally intensive and complex to use, which makes optimization and decision-making challenging. This dissertation therefore proposes a hybrid physics-informed machine learning optimization system for waterflooding management and injection allocation optimization, developed and validated using data from the EMN sub-field of the El Merk oil field in the Berkine Basin in Algeria. The system integrates a Capacitance-Resistance Model (CRM) for connectivity, a Long Short-Term Memory (LSTM) model for forecasting fractional flow, a Physics-Informed Neural Network (PINN) for solving the heterogeneous Buckley–Leverett equation, and a constrained Sequential Least Squares Programming (SLSQP) optimizer for injection allocation. This system was developed using 8460 well-test observations for the machine learning models and 27 574 allocated observations for the CRM model, drawn from six producers and two injectors over 156 months. The LSTM model achieved a coefficient of determination (R²) of 0.944, while the PINN delivered an overall R² of 0.978. The optimizer produced the best possible injection rates, leading to an estimated incremental oil volume uplift of 222 285 STB per year. This hybrid system is more efficient, faster, and simpler than traditional methods, demonstrating its potential to optimize waterflooding and improve oil recovery