Hybrid Machine Learning and Meta-Heuristic Optimization for Well Placement and Operating Conditions in Carbonate Rocks: Low-Salinity Water Injection

Technical Summary: Hybrid Machine Learning and Meta-Heuristic Optimization for Well Placement and Operating Conditions in Carbonate Rocks

Problem Statement
Low-salinity water injection (LSWI) is a recognized enhanced oil recovery (EOR) technique, particularly effective in heterogeneous carbonate reservoirs where it improves sweep efficiency through wettability alteration and ion exchange. However, optimizing LSWI processes is computationally prohibitive. Accurate modeling requires compositional reservoir simulations to capture complex rock-fluid interactions, which are time-intensive (taking several hours per scenario). Consequently, direct optimization of well placement and operating conditions using full-physics models is impractical for large decision spaces. Traditional trial-and-error approaches are inefficient, and while previous studies have optimized LSWI parameters using algorithms like DECE, they have largely relied on physics-based models or focused on sandstone reservoirs. There is a gap in the literature regarding the simultaneous optimization of well locations and operating conditions in carbonate reservoirs using a data-driven framework that couples surrogate models with meta-heuristic algorithms.

Methodology
The study proposes a data-driven framework that integrates Artificial Neural Network (ANN) surrogate models with multi-objective meta-heuristic optimization algorithms (Genetic Algorithm, Particle Swarm Optimization, and Gray Wolf Optimization). The workflow proceeds as follows:

1. Reservoir Modeling and Simulation: A sector model of a heterogeneous carbonate reservoir (6020 ft × 6020 ft × 100 ft) was constructed using a compositional simulator with the Peng-Robinson Equation of State (PR-EOS). The model incorporated clay minerals, quartz, calcite, and dolomite, with heterogeneous distributions of porosity, permeability, and cation exchange capacity (CEC). Relative permeability curves were interpolated based on ion exchange mechanisms between high-salinity and low-salinity conditions.
2. Data Generation: To train the surrogate models, experimental design tables were generated using the Latin Hypercube method.
   • Well Placement Phase: 665 simulation runs were conducted to generate data linking well block coordinates (mapped to local porosity, permeability, and CEC) to target parameters: oil recovery factor (RF), water cut (WC), and Net Present Value (NPV).
   • Operating Conditions Phase: A second set of simulations was run using the optimal well locations to generate data linking operating variables (injection rate, production rate, bottom-hole pressure, and perforation status) to target parameters (RF, cumulative water production, and NPV).
3. Surrogate Model Development: Two distinct ANNs were developed:
   • A three-hidden-layer network (54-25-13 neurons) for well placement optimization.
   • A single-hidden-layer network (42 neurons) for operating condition optimization.
     Both were trained using Bayesian regularization, with 70-80% of data used for training and the remainder for testing.
4. Optimization: The trained ANNs served as fast proxies for the full-physics simulator. They were coupled with GA, PSO, and GWO algorithms to identify Pareto-optimal solutions that maximize NPV and RF while minimizing water production.
5. Validation: Selected solutions from the Pareto fronts were validated by running them through the full-physics simulator to calculate relative errors.

Key Results
The study demonstrates significant improvements in both computational efficiency and reservoir performance:

• Computational Efficiency: The surrogate model reduced the evaluation time for each scenario from approximately 3 hours (full-physics) to about 1 second. The entire optimization process for each algorithm took roughly 50 minutes, compared to several months required for full-physics optimization.
• Model Accuracy: The ANN models achieved high predictive accuracy, with R-squared values of 0.963 (training) and 0.942 (testing) for well placement, and over 0.9 for operating conditions. The mean square errors were 0.003 and 0.0027, respectively.
• Performance Improvements:
  • Well Placement Optimization: Using the Genetic Algorithm (which yielded the lowest error), optimizing well locations increased the NPV by $70 million and improved the oil recovery factor by 6% (from 50.1% to 56.1%). Optimal wells were placed in blocks with high porosity, permeability, and CEC.
  • Operating Conditions Optimization: Using the Gray Wolf Optimization algorithm (selected for lowest error in this phase), optimizing operating parameters (pressure, rates, perforation status) further increased the NPV by $160 million and raised the oil recovery factor to 66.3%.
  • Combined Optimization: The simultaneous optimization of well placement and operating conditions resulted in a total NPV increase of $230 million, a 16.2% increase in oil recovery, and a reduction of 290 MMSTB in cumulative water production compared to the base case.

Significance and Claims
The authors claim that this study provides a computationally efficient alternative to full-physics optimization for enhancing LSWI performance in carbonate reservoirs. The primary contribution is the successful integration of ANN-based surrogate models with multi-objective meta-heuristic algorithms to simultaneously optimize well locations and operating conditions—a combination not previously investigated for LSWI.

The paper asserts that this framework allows for the exploration of large, complex decision spaces that would otherwise be infeasible due to computational constraints. By reducing simulation time from hours to seconds without sacrificing predictive accuracy, the proposed method supports faster, data-driven decision-making for field-scale EOR projects. The study concludes that while the current framework assumes static petrophysical properties, it establishes a robust baseline for optimizing LSWI, with future work suggested to integrate petrophysical uncertainty via Monte Carlo simulations.