Physics-informed machine learning with differentiable programming for heterogeneous underground reservoir pressure...

Area of Science:

• Geosciences
• Computational Science
• Machine Learning

Background:

• Subsurface reservoir pressure management is critical for CO2 sequestration and wastewater injection.
• Subsurface heterogeneity and parametric uncertainty complicate accurate pressure predictions.
• High-fidelity physics-based models are computationally intensive for managing reservoir pressures.

Purpose of the Study:

• To develop a computationally efficient framework for managing subsurface reservoir pressures.
• To integrate differentiable programming and machine learning for real-time reservoir pressure control.
• To accurately predict fluid extraction rates that prevent reservoir over-pressurization.

Main Methods:

• Utilized a differentiable programming framework (DPFEHM) with full-physics, two-point flux finite volume discretization.
• Employed physics-informed machine learning with convolutional neural networks to learn extraction rates based on permeability fields.
• Conducted hyperparameter searches and executed training/testing scenarios for model evaluation.

Main Results:

• Developed a physics-informed machine learning framework capable of managing reservoir pressures.
• Achieved a simulator that is 400,000 times faster than the underlying physics-based simulator.
• Enabled near real-time analysis and robust uncertainty quantification for reservoir management.

Conclusions:

• Physics-informed machine learning offers a viable and highly efficient solution for subsurface reservoir pressure management.
• The developed framework addresses the computational challenges posed by subsurface heterogeneity and uncertainty.
• This approach facilitates safer and more effective CO2 sequestration and wastewater injection operations.