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

This paper presents a physics-informed convolutional neural network framework that accurately predicts pore-scale velocity fields in complex porous media by integrating physical constraints into the training process, thereby enabling significant acceleration of Lattice-Boltzmann simulations through improved initial conditions.

The Gist

Imagine you are trying to predict how water flows through a sponge. In the real world, sponges have tiny, twisting, and irregular holes. To calculate exactly how the water moves through every single twist and turn using traditional math, you need a supercomputer and a lot of time. It's like trying to map every single grain of sand on a beach by hand; it's accurate, but painfully slow.

This paper introduces a new way to do this using Artificial Intelligence (AI). Think of the AI as a "super-observer" that learns to guess the water's path just by looking at a picture of the sponge's holes, without needing to do the heavy math every time.

Here is a breakdown of how they did it and what they found, using simple analogies:

1. The Problem: The "Slow Math" vs. The "Fast Guess"

Traditionally, scientists use a method called Lattice Boltzmann Method (LBM) to simulate fluid flow. Imagine this as a very careful, slow-motion video game where the computer calculates the movement of billions of tiny water particles one by one. It's accurate, but it takes a long time to run, especially for complex sponges.

The authors wanted to train a Convolutional Neural Network (CNN)—a type of AI good at recognizing patterns in images—to act as a "shortcut." They wanted the AI to look at a picture of the sponge and instantly "paint" the picture of how the water would flow through it.

2. The Training: Teaching the AI with "Rules"

You can't just show the AI pictures and let it guess randomly. If you do, it might draw water flowing through the solid parts of the sponge, which is physically impossible.

To fix this, the authors gave the AI a special scorecard (Loss Function) with four specific rules to follow, much like a coach correcting a student:

• The "No-Go" Zone Rule: If the AI predicts water flowing inside a solid rock or obstacle, it gets a big penalty. (Imagine a teacher saying, "Water can't walk through walls!")
• The "No-Spilling" Rule: The water must be incompressible (it can't just disappear or appear out of nowhere). The AI is penalized if the math doesn't balance out.
• The "Seamless Wrap" Rule: Since the sponge samples are treated as if they wrap around like a video game map (periodic boundaries), the flow on the left edge must match the flow on the right edge. The AI is penalized if the flow looks broken at the edges.
• The "Twistiness" Rule: The AI must predict the correct overall "tortuosity" (how twisty and long the path is). If the path looks too straight or too crazy compared to reality, it loses points.

By combining these rules with the actual answer (the slow, accurate LBM simulation), the AI learned to make guesses that were not only fast but also physically correct.

3. The Results: The "Best Student"

The researchers tested many different AI architectures (different "brain" designs). They found that a specific design called ResNet-101 was the best student.

• Accuracy: It could predict the water flow with incredible precision, matching the slow, expensive computer simulations almost perfectly.
• Speed: While the traditional method took hundreds of milliseconds, the AI could make a prediction in just 5 milliseconds on a graphics card. That's like going from walking to a sprint.

4. The "Out-of-Distribution" Test: Can it Handle New Sponges?

A smart AI shouldn't just memorize the training pictures; it should understand the concept of flow. The researchers tested the AI on sponges it had never seen before:

• Different Shapes: They used sponges made of squares and circles instead of the wavy lines the AI was trained on. The AI still worked well, though it struggled slightly more with sharp squares than round circles.
• Different Densities: They tested sponges that were very dense (few holes). The AI did well on moderately dense sponges but started to get confused when the sponge was extremely dense (near the point where water can't flow through at all).
• Real-World Sponges: They even tested it on real lithium-ion battery electrodes (scanned from real life). The AI handled these messy, real-world structures surprisingly well.

5. The "Superpower" Application: The Warm Start

The most practical trick they discovered is using the AI to speed up the slow computer simulations.

• The Cold Start: Usually, to run a simulation, you start with zero water movement and wait for it to settle. This takes a long time.
• The Warm Start: The researchers let the AI make a quick, "rough guess" of the flow first. They fed this guess into the slow computer simulation as a starting point.
• The Result: Because the simulation started with a good guess instead of zero, it converged (finished) 50% faster in half of the cases. In 90% of cases, it was faster than starting from scratch.

Summary

The paper presents a system where an AI learns to predict fluid flow through porous materials by looking at the shape of the holes. By teaching the AI strict physical rules (like "water can't go through rocks"), they created a tool that is:

1. Extremely fast (milliseconds vs. seconds).
2. Physically accurate (it respects the laws of physics).
3. Versatile (it works on new shapes and even real-world materials).
4. A booster (it can speed up traditional simulations by giving them a "head start").

The authors conclude that while the AI isn't perfect for every single extreme case (like extremely dense sponges), it is a powerful new tool for understanding how fluids move through complex materials.

Technical Summary

Technical Summary: Physics-Informed Convolutional Neural Networks for Fluid Flow Through Porous Media

Problem Statement
Accurate simulation of fluid flow in porous media is hindered by the geometric complexity of pore spaces and the high computational cost of solving the Navier–Stokes equations. Traditional numerical solvers, such as the Lattice Boltzmann Method (LBM), require careful mesh construction and often exhibit slow convergence, particularly in diffusive momentum transport regimes found in intricate solid boundaries. While deep learning has shown promise in predicting physical fields, applying it to complex, variable-porosity porous media remains challenging due to the need for models that respect physical constraints (e.g., incompressibility, no-slip conditions) and generalize beyond specific training distributions.

Methodology
The authors propose a physics-informed convolutional neural network (CNN) framework to predict pore-scale velocity fields directly from binary images of porous sample geometries.

• Data Generation: A dataset of 10,000 synthetic 2D porous structures (256×256 pixels) was generated using superimposed standing sinusoidal waves with random phases and directions. Porosity ($\phi$) was uniformly sampled between 0.70 and 0.95. Ground-truth velocity fields were generated using the Lattice Boltzmann Method (LBM) under periodic boundary conditions and low Reynolds numbers.
• Architecture: The model utilizes a U-Net architecture with an encoder-decoder structure and skip connections to preserve spatial details. Various convolutional backbones (VGG, ResNet, DenseNet, ConvNeXt, EfficientNet, MobileNet) were evaluated.
• Physics-Informed Loss Function: A custom loss function ($L$) was developed to enforce physical consistency, composed of five terms:
  1. Velocity Loss ($L_{vel}$): Mean squared error between predicted and reference velocity fields.
  2. Obstacle Loss ($L_{obstacle}$): An $L_1$ norm penalty enforcing zero velocity within solid regions ($B$).
  3. Divergence Loss ($L_{div}$): Penalizes non-zero divergence ($\nabla \cdot u$) in the fluid pore space ($P$) to enforce incompressibility.
  4. Periodicity Loss ($L_{perio}$): Encourages consistency under periodic translations by penalizing discrepancies between predictions for a translated structure and the translated prediction of the original structure.
  5. Tortuosity Loss ($L_{tort}$): Penalizes the squared difference between predicted and reference tortuosity ($\tau$), a global macroscopic property.
• Training Strategy: A two-stage training protocol was employed. A "pilot" model was first trained with only velocity loss, then fine-tuned using the full loss function with optimized weights ($\alpha=5, \beta=1, \gamma=0.1, \delta=0.01$). Data augmentation (vertical flipping, random rolling) was applied to enhance generalization.

Key Results

• Architecture Performance: Among the tested backbones, ResNet-101 achieved the best performance, yielding the lowest Root Mean Squared Error (RMSE) for velocity predictions ($5.1 \times 10^{-3}$) and high coefficients of determination ($R^2$) for derived properties: 0.983 for tortuosity and 0.997 for permeability.
• Loss Function Analysis: Systematic ablation studies demonstrated that while velocity loss alone minimized MSE, it failed to suppress spurious flow in solids or predict tortuosity accurately. The inclusion of obstacle, divergence, periodicity, and tortuosity terms significantly improved physical plausibility and global property prediction, with the full loss function achieving an $R^2$ of 0.983 for tortuosity.
• Generalization: The model demonstrated robust generalization to out-of-distribution samples without retraining:
  • Geometry: It successfully predicted flows in structures with square and mixed obstacles, though accuracy decreased slightly as the proportion of square obstacles increased.
  • Boundary Conditions: The model generalized to "pipe-like" geometries with confined boundaries, correctly reconstructing unidirectional flow profiles despite never seeing such constraints during training.
  • Porosity: Performance remained high for porosities within and slightly below the training range (0.6–0.8). However, accuracy degraded significantly for very low porosities (0.5–0.6) near the percolation threshold, where transport is dominated by narrow channels and connectivity transitions.
  • Real-World Data: When tested on 2D slices of real Li-O2 electrode microstructures (highly irregular, heterogeneous), the model maintained reasonable accuracy for permeability ($R^2=0.986$), though tortuosity prediction was less accurate ($R^2=0.798$) due to geometric complexity.
• LBM Initialization: A practical application was demonstrated where the CNN-predicted velocity field served as a "warm start" for the LBM solver. This approach accelerated LBM convergence in over 90% of cases, reducing the median number of iterations by approximately 50% compared to a "cold start" (zero velocity).
• Efficiency: Inference on a GPU takes approximately 5 ms per structure, compared to an average of 560 ms for a full LBM simulation on a CPU.

Significance and Claims
The authors position this work not as the introduction of a new learning paradigm, but as a specific formulation and rigorous evaluation of a problem-specific physics-informed loss function for pore-scale flow prediction. The primary contributions are:

1. The development of a complete pipeline from synthetic data generation to CNN training for velocity field prediction.
2. The construction and systematic analysis of a multi-term loss function that balances local velocity accuracy with global physical constraints (impermeability, incompressibility, periodicity, and tortuosity).
3. The demonstration that such models can generalize to varying obstacle shapes, boundary conditions, and porosities, provided the fundamental flow characteristics remain similar.
4. The validation of CNN predictions as effective initial conditions for traditional solvers, significantly reducing computational time for complex boundary problems.

The paper acknowledges limitations, noting that the current implementation is restricted to 2D, fixed-resolution geometries and struggles with extreme low-porosity regimes near percolation thresholds. The authors suggest that future work could extend this to 3D systems and explore transfer learning for experimentally obtained microstructures.