Accelerating Physics-Based Simulations Using End-to-End Neural Network Proxies: An Application in Oil Reservoir...

Area of Science:

• Petroleum Engineering
• Artificial Intelligence
• Computational Science

Background:

• Traditional physics-based PDE solvers for oil reservoir simulation are computationally intensive, limiting optimization speed.
• Accelerating reservoir simulations is crucial for efficient oil field development and decision-making.

Purpose of the Study:

• To develop and evaluate a deep learning proxy model for significantly accelerating oil reservoir simulations.
• To demonstrate the effectiveness of an end-to-end black-box modeling approach for reservoir simulation tasks.
• To provide a benchmark for AI-driven optimization in oil and gas exploration.

Main Methods:

• Developed a proxy model utilizing deep learning techniques to emulate physics-based PDE solvers.
• Modeled the reservoir simulator as an end-to-end black box for comprehensive simulation.
• Conducted experimental evaluations on a publicly available reservoir model with varied well locations and geological realizations.

Main Results:

• Achieved a speedup of over 2000X compared to industry-standard physics-based solvers.
• Maintained an average sequence error of approximately 10% relative to the physics-based simulator.
• Outperformed baseline methods, including upscaling, by two orders of magnitude.

Conclusions:

• The deep learning proxy model offers a highly efficient and accurate solution for accelerating oil reservoir simulations.
• The end-to-end black-box approach demonstrates significant potential for practical applications in oil field development optimization.
• The domain-agnostic architecture of the model allows for potential extension to diverse applications beyond the oil and gas industry.