Abstract
Geological Carbon Capture and Storage (CCS) requires the strategic placement of deep injection wells to maximize $CO_2$ storage while maintaining reservoir integrity. This study develops a high-performance optimization tool combining the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) with the TOUGH2-MP simulator. By implementing a two-level parallel computing approach on the Oakforest-PACS supercomputer, the researchers significantly reduced the computational time for finding optimal well configurations from months to weeks, demonstrating the feasibility of real-scale optimization in practical timeframes.
Introduction
Mitigating climate change through CCS involves injecting millions of tons of $CO_2$ into deep underground formations. A primary engineering challenge is determining the optimal placement of injection wells to maximize capacity and prevent excessive pressure buildup that could fracture rock or cause leaks. Traditional optimization methods are often computationally prohibitive due to the need for thousands of complex 3D fluid dynamics simulations. This research addresses the bottleneck of computational time by leveraging massive parallel computing and metaheuristic algorithms to automate and accelerate the design process.
Methods
The proposed tool utilizes a "double-layered" parallel computing architecture implemented on the Oakforest-PACS supercomputer:
- Optimization Algorithm (CMA-ES): The Covariance Matrix Adaptation Evolution Strategy was selected for its parameter-free nature and efficiency in exploring complex search spaces. It works by evolving a population of potential well layouts based on the performance of previous generations.
- Simulation Engine (TOUGH2-MP): To evaluate each layout, the tool uses a parallelized version of the industry-standard TOUGH2 simulator, which models the 3D flow of water and supercritical $CO_2$ in porous media.
- Two-Level Parallelism: The core innovation lies in parallelizing both the fluid simulations and the optimization process itself. By evaluating 20 different well-placement candidates simultaneously, with each simulation utilizing 64 CPU cores, the system utilizes 1,280 cores at once to explore the search space rapidly.
The study evaluated the tool using a realistic 3D reservoir model (8 km x 12 km x 0.15 km) featuring heterogeneous porosity and permeability based on actual seismic data.
Key Insights and Results
The optimization process demonstrated a dynamic search strategy. Initially, the algorithm explores a wide area (high variance) to identify potential regions of interest. As it identifies "sweet spots" with high permeability and low initial pressure, it narrows its focus to fine-tune the coordinates.
Performance and Capacity Gains
- Massive Speedup: For a 3-well placement problem, the traditional approach (parallelizing only the simulation) would have taken roughly 105 days. The new two-level parallel approach completed the task in just 13 days.
- Increased Efficiency: The optimized layout increased the total $CO_2$ injection amount by 10% compared to a manual "best guess" by an engineer.
- Pressure Management: The tool highlighted that adding "pressure relief" wells to pump out brine can be more effective than simply adding more injection wells. In specific cases, this strategy increased $CO_2$ storage capacity by 30%.
Discussion and Conclusion
While the tool offers massive speedups, it currently necessitates access to supercomputing resources or large cloud clusters to be practical. Future work will focus on improving load balancing to ensure that CPU cores do not sit idle while waiting for the slowest simulations in a batch to finish. Despite these requirements, the development of this tool marks a significant step toward making large-scale CCS site design practical and efficient for real-world applications.
Reference: Miyagi, A., Yamamoto, H., & Akimoto, Y. (2021). Development of an Optimization Tool for Well Placement Optimization in Geological Carbon Dioxide Sequestration by Metaheuristics. Journal of JSCE A2 (Applied Mechanics), Vol. 77, No. 1, 21-34.
Summary updated