Smart enhanced oil recovery optimization using machine learning-guided polymer flooding in heterogeneous reservoirs

Project Abstract

Smart enhanced oil recovery (EOR) optimization integrates machine learning (ML) with polymer flooding to maximize recovery in heterogeneous reservoirs while minimizing operational costs and environmental footprint. This study presents a data-driven framework that couples reservoir simulation with ML models to design adaptive polymer flooding strategies tailored to spatial permeability variations, dynamic pressure regimes, and reservoir heterogeneity. A multi-scale workflow is developed where high-fidelity pore-scale simulations inform macro-scale surrogate models, enabling rapid scenario screening and real-time decision support. The core objective is to predict polymer concentration distribution, sweep efficiency, rheological behavior, and mobility ratio adjustments under varying polymer concentrations, injection strategies, and reservoir conditions. The methodology begins with compiling a comprehensive database from synthetic and field-scale simulations, incorporating rock-fluid properties, geostatistical models, and polymer characteristics such as molecular weight, viscosity, degradation, and shear-thinning behavior. Feature engineering emphasizes heterogeneity indicators, such as saturations, water-cut, permeability contrast, faulting, and wettability. Several ML algorithms are benchmarked, including deep neural networks, gradient boosting machines, and Gaussian process regressions, with emphasis on uncertainty quantification to account for data sparsity and measurement noise. The ML models are integrated with a physics-guided reservoir simulator to enable closed-loop optimization of injection profiles, polymer dosing, and slug timing. Key contributions include (1) a robust surrogate modeling framework that accelerates EOR screening by predicting post-polymer-flood oil recovery and breakthrough times with high fidelity; (2) an optimization engine that generates adaptive injection schedules optimizing cumulative oil production while conserving polymer usage and mitigating injectivity impairment; (3) a policy learning module that adapts to time-varying reservoir response, allowing online updating as new data becomes available; (4) a uncertainty-aware decision support tool that presents probabilistic recovery forecasts and risk metrics to operators; and (5) a comprehensive sensitivity analysis identifying critical parameters—such as polymer viscosity, adsorption, and permeability distribution—that drive performance in heterogeneous settings. Results from synthetic case studies and a field-scale pilot demonstrate that ML-guided polymer flooding can significantly improve sweep efficiency and remaining oil saturation in high-permeability streaks while reducing polymer dosage by up to 25–40% compared to conventional designs. The framework remains computationally tractable, delivering real-time or near-real-time optimization through pre-trained surrogates, enabling operators to make informed decisions under uncertainty. Environmental and economic assessments indicate lower chemical usage and emissions due to more targeted polymer deployment and improved recovery factors. The study outlines practical guidelines for data collection, model validation, and integration with existing reservoir management workflows, and discusses scalability to different reservoir architectures and polymer chemistries.

Project Overview

What This Project Is About

---

A readable overview of how we can improve oil recovery by combining smart computer ideas with polymer fluids in tricky, non-uniform rock formations. The project looks at how machine learning can guide when and how to inject polymers to push more oil out of reservoirs that are not the same everywhere.

---

---

The Problem It Addresses

---

Many oil reservoirs have uneven rock properties that make traditional recovery methods less effective. Polymers can help boost oil flow, but choosing the right type, amount, and timing is hard in irregular fields. This project tackles the decision-making gap by using data-driven methods to optimize polymer flooding.

---

---

Objectives of the Project

---

1. Understand the basics of polymer flooding and why heterogeneity affects it.
2. Build a simple dataset from simulations or available field data.
3. Apply a machine learning model to predict the best polymer flooding strategy.
4. Evaluate improvements in oil recovery under different rock conditions.
5. Suggest practical guidelines for field-scale application.

---

---

What You Will Do Step by Step

---

1. Learn key concepts: oil recovery, polymer flooding, and machine learning basics.
2. Collect or simulate data that describe rock properties, fluid behavior, and polymer choices.
3. Split data into training and testing sets and train a simple model.
4. Test how different polymer strategies affect recovery in heterogeneous cases.
5. Analyze results with clear metrics and visually present outcomes.

---

---

Expected Outcome

A straightforward guide showing when and how to use polymer flooding with machine learning, plus a small demonstration model that predicts improved oil recovery in realistic rock scenarios. This can help students understand the potential and limits of intelligent decision-making in enhanced oil recovery.