Physics-informed convolutional neural networks for fluid flow through porous media

Area of Science:

• Computational fluid dynamics
• Porous media flow simulation
• Artificial intelligence in scientific computing

Background:

• Simulating fluid flow in porous media is computationally intensive due to complex geometries and Navier-Stokes equations.
• Traditional meshing methods are slow, require manual input, and struggle with intricate pore structures.
• Existing methods face challenges in efficiency, especially for repeated simulations in complex boundary problems.

Purpose of the Study:

• To develop a novel neural network framework for direct prediction of pore-scale velocity fields from geometry.
• To enhance the efficiency and speed of fluid dynamics simulations in porous media.
• To investigate the generalization capabilities and practical applications of the AI framework.

Main Methods:

• A convolutional encoder-decoder neural network with skip connections was employed to capture fine-grained structural details.
• A custom loss function incorporating incompressibility, no-flow, periodicity, and tortuosity constraints ensured physical consistency.
• Extensive analysis of loss term weights and architectural variants, including generalization tests on diverse datasets.

Main Results:

• The neural network framework accurately predicts pore-scale velocity fields directly from sample geometry.
• Physical consistency was maintained through a carefully designed multi-term loss function.
• Network predictions significantly accelerate Lattice-Boltzmann simulation convergence, improving speed in over 90% of tested cases.

Conclusions:

• The proposed neural network offers an efficient alternative for simulating fluid flow in porous media.
• This AI-driven approach demonstrates robust generalization across various boundary conditions and geometries.
• Utilizing network-generated velocity fields as initializations substantially speeds up Lattice-Boltzmann solvers for complex flow problems.