An Adaptive Machine Learning Framework for Fluid Flow in Dual-Network Porous Media

This paper presents a physics-informed neural network framework with adaptive weighting and dynamic collocation strategies to enable rapid, mesh-free forward and inverse modeling of fluid flow in dual-porosity/permeability porous media, offering robust parameter identification and accurate handling of complex geometries and solution discontinuities.

The Gist

The Big Picture: Two Worlds in One Rock

Imagine a sponge, but not just any sponge. Imagine a sponge made of two different materials glued together:

1. The Big Holes: Think of large cracks and tunnels (like a cave system). Water flows through these very fast.
2. The Tiny Pores: Think of the tiny, dense holes inside the rock material itself. Water moves through these very slowly, like it's stuck in molasses.

In the real world, rocks (like shale for oil or soil for groundwater) often look like this. They have a "Dual Network." The water in the big holes and the water in the tiny pores are constantly swapping places.

The Problem:
Scientists need to predict how water (or oil) moves through these rocks.

• Old Way (The Grid): Traditional computer models try to draw a grid over the rock, like graph paper. If the rock has sudden changes (like a layer of clay next to a layer of sand), the grid gets confused. It either creates "ghost ripples" (errors) or needs to be so incredibly detailed that the computer takes days to crunch the numbers.
• The New Way (The Paper's Solution): The authors created a new tool using Artificial Intelligence (AI) called a Physics-Informed Neural Network (PINN).

---

The Solution: The "Smart Detective" AI

Instead of drawing a grid, the AI acts like a Smart Detective. Here is how it works, broken down into simple steps:

1. The "Shared Brain" (The Architecture)

Usually, if you want an AI to predict two things (like pressure in the big holes AND pressure in the tiny pores), you might build two separate brains. But that's inefficient.

• The Analogy: Imagine a master chef (the Shared Trunk) who learns the general rules of cooking (the physics of fluid flow). Then, they have two specialized assistants (the Slim Heads). One assistant only cares about the sauce (Macro-pressure), and the other only cares about the garnish (Micro-pressure).
• Why it helps: They share the same knowledge base. If the sauce changes, the garnish assistant instantly knows why, because they share the same brain. This keeps the math consistent and prevents the AI from getting confused.

2. The "Smart Spotlight" (Adaptive Sampling)

When the AI tries to learn, it doesn't look at the whole rock equally.

• The Analogy: Imagine you are trying to learn a map of a city. You don't need to study the empty fields in the middle of nowhere. You need to focus on the busy intersections and the confusing one-way streets.
• How it works: The AI looks at where it is making mistakes (where the "residual" is high). It then says, "Hey, I'm confused here! Let's put more 'spotlights' (data points) in this specific area to learn better." It ignores the easy parts and zooms in on the hard parts.

3. The "Balanced Scale" (Adaptive Weighting)

The AI has to satisfy two rules:

1. The Physics Rule: The water must obey the laws of physics (conservation of mass, etc.).
2. The Boundary Rule: The water must match what we know happens at the edges (e.g., "No water flows out the back door").

• The Analogy: Imagine a student taking a test. They have to answer Math questions (Physics) and History questions (Boundaries). Sometimes the Math teacher is grading harder than the History teacher. If the student focuses too much on Math, they fail History.
• How it works: The AI automatically adjusts how much it cares about each rule. If it's struggling with the Physics rules, it gives them more "weight" in its brain. If it's struggling with the boundaries, it shifts focus there. It balances itself dynamically so it doesn't get stuck.

---

Why is this a Big Deal?

The paper shows that this new AI framework is a "Superhero" compared to old methods for three main reasons:

1. No Grid Needed (Mesh-Free):

   • Old Way: You have to build a perfect Lego structure (mesh) to fit the rock. If the rock has a weird shape, the Legos don't fit, and you have to rebuild everything.
   • New Way: The AI just floats over the shape. It doesn't care if the rock is a perfect square or a jagged, broken piece of shale. It works on any shape instantly.

2. Handles "Cliffs" Perfectly:

   • Old Way: If you have a layer of rock with high permeability next to a layer with low permeability, the old math often creates "wobbly lines" or fake ripples (spurious oscillations) at the boundary. It's like trying to draw a sharp corner with a soft pencil; it smudges.
   • New Way: The AI draws a razor-sharp line. It captures the sudden jump in speed perfectly, just like a discontinuous Galerkin method (a very advanced, expensive math technique) but without the headache.

3. It Can Work Backwards (Inverse Problems):

   • The Scenario: Usually, we know the rock's properties and want to know the flow. But in the real world (like mining or oil drilling), we often know the flow (we see water coming out of a well) but we don't know the rock's hidden properties (like how fast water swaps between the big and small pores).
   • The Magic: This AI can work backward. You tell it, "Here is the water coming out of the well." It figures out, "Ah, the swapping speed must be this value." It solves the mystery of the hidden rock properties using the data we can actually measure.

The Bottom Line

The authors have built a smart, self-correcting AI that learns how fluids move through complex, dual-layered rocks. It doesn't need a grid, it focuses on the hard parts, it balances its own learning rules, and it can even solve mysteries about hidden rock properties.

It's like upgrading from a slow, rigid calculator to a flexible, intuitive assistant that understands the story of the fluid flow, not just the numbers.

Technical Summary

1. Problem Statement

The paper addresses the challenge of modeling fluid flow in dual-porosity/permeability (DPP) systems. These systems, common in fractured rocks, tight shales, and engineered composites, consist of two interconnected pore networks (macro-pores and micro-pores) with distinct properties and mass exchange between them.

Key Challenges:

• Mathematical Complexity: Unlike classical Darcy flow (single network), DPP models require solving a system of four coupled partial differential equations (PDEs) involving two pressure fields and two velocity fields. These equations cannot be reduced to a single pressure-only formulation.
• Numerical Instability: Traditional Finite Element Methods (FEM) struggle with DPP models due to the Ladyzhenskaya–Babuška–Brezzi (LBB) stability condition. Standard equal-order interpolation often leads to spurious oscillations (Gibbs phenomenon) and requires complex stabilization techniques or Discontinuous Galerkin (DG) methods.
• Inverse Problems: Critical parameters, such as the inter-porosity mass transfer coefficient ($\beta$), are difficult to measure directly but are essential for accurate forecasting and resource extraction (e.g., hydrocarbon recovery, mineral exploration).

2. Methodology: Adaptive Physics-Informed Neural Network (PINN)

The authors propose a novel, mesh-free PINN framework specifically designed to handle the coupled nature of DPP systems. The methodology integrates three core algorithmic advancements:

A. Shared-Trunk with Slim-Head Architecture

To address the strong coupling between the four field variables ($p^{(1)}, p^{(2)}, u^{(1)}, u^{(2)}$), the authors avoid training separate networks for each variable. Instead, they employ a Multi-Task Learning (MTL) architecture:

• Shared Trunk: A deep neural network backbone processes the input coordinates (enhanced by Fourier feature encoding) to learn a unified latent representation of the physical domain.
• Slim Heads: Lightweight, task-specific output layers branch from the trunk to predict individual fields (pressure and velocity for both networks).
• Benefit: This design enforces physical consistency, reduces parameter redundancy, and improves data efficiency by leveraging shared representations of the underlying physics.

B. Adaptive Weighting Strategy

Training PINNs often suffers from imbalanced convergence between different loss components (PDE residuals vs. Boundary Conditions). The paper introduces a normalized learning speed-based adaptive weighting scheme:

• The algorithm tracks the fractional loss reduction (learning speed) of each task.
• Weights are dynamically adjusted to balance the optimization, ensuring that slower-converging tasks (e.g., specific boundary conditions or coupling terms) receive higher priority, preventing the network from ignoring critical constraints.

C. Residual-Based Adaptive Refinement (RAR)

To resolve sharp gradients and discontinuities typical in layered porous media, the framework employs adaptive sampling:

• During training, the network evaluates residuals at a pool of candidate points.
• Points with the highest residuals (indicating high error) are selected and added to the training collocation set.
• This focuses computational resources on regions requiring higher resolution, such as interfaces between layers with contrasting permeabilities.

D. Mathematical Formulation

The framework solves the DPP equations in their mixed form (simultaneously solving for pressure and velocity). The loss function is a weighted sum of:

1. PDE Residuals: Enforcing momentum balance and mass conservation (including the inter-porosity transfer term $\beta$).
2. Boundary Conditions: Enforcing Dirichlet (pressure) and Neumann (velocity) conditions.
3. Observation Misfit (for Inverse Problems): Minimizing the difference between predicted and observed data (e.g., total flux).

3. Key Contributions

1. First PINN Framework for DPP: This is the first systematic integration of DPP couplings into a neural network architecture, moving beyond single-field formulations.
2. Stability Without Stabilization: The framework naturally satisfies the LBB condition and captures discontinuities without the spurious oscillations seen in classical FEM or the need for DG stabilization terms.
3. Robust Inverse Modeling: The framework successfully identifies difficult-to-measure parameters (like $\beta$) by jointly optimizing network weights and physical parameters against observational data.
4. Theoretical Convergence Analysis: The authors provide a rigorous mathematical proof of the existence and uniqueness of the solution to the infinite-dimensional least-squares problem associated with the DPP model, establishing the well-posedness of the framework.

4. Numerical Results

The framework was validated against analytical solutions and high-fidelity Finite Element/Discontinuous Galerkin benchmarks across several scenarios:

• 1D Benchmark Problems:
  • Pressure-Driven Flow: Results matched stabilized mixed FEM solutions perfectly, demonstrating the ability to handle standard flow scenarios.
  • Mixed Boundary Conditions: The model correctly predicted non-uniform micro-velocities driven by internal mass exchange, even when the micro-network had no boundary outflow—a phenomenon classical Darcy models fail to capture.
• 2D Radial Symmetry (Candle Filter): The model accurately reproduced radial pressure and velocity profiles, preserving symmetry and capturing internal flow exchange in a micro-pore network with zero boundary flux.
• 2D Layered Media (Heterogeneity): The framework handled abrupt permeability changes across five distinct layers. It captured velocity discontinuities at interfaces with high accuracy, outperforming continuous FEM and matching DG results without requiring DG-specific formulations.
• 2D Strip Footing: A complex geotechnical benchmark with mixed boundary conditions showed excellent agreement with reference solutions.
• Inverse Problem (Parameter Identification): In a 2D pressure-driven problem, the framework successfully recovered the mass transfer coefficient ($\beta$) from total flux observations. The recovered $\beta$–flux relationship matched the forward FEM solution, validating its utility for data assimilation.

Performance:

• Training times were efficient (e.g., ~6–9 minutes on an NVIDIA T4 GPU).
• The adaptive strategies (weighting and sampling) significantly improved convergence rates and solution accuracy compared to static training approaches.

5. Significance and Impact

• Mesh-Free Advantage: The approach eliminates the need for complex mesh generation, making it ideal for porous media with irregular geometries or evolving fracture networks.
• Handling Discontinuities: By avoiding the spectral bias of standard neural networks through adaptive sampling and shared architectures, the method accurately resolves sharp gradients and layer interfaces where classical methods often fail or oscillate.
• Inverse Analysis Capability: The framework bridges the gap between physical modeling and data availability, enabling the estimation of critical subsurface parameters that are otherwise inaccessible, which is crucial for resource management and environmental monitoring.
• Future Directions: The authors suggest extending the framework to nonlinear flow regimes (e.g., Darcy–Forchheimer) and coupled poromechanics (Biot's equations), further expanding the applicability of physics-informed machine learning in geotechnical and petroleum engineering.

In conclusion, this paper presents a robust, mathematically grounded, and computationally efficient framework that overcomes the limitations of traditional numerical methods for dual-porosity flow, offering a powerful tool for both forward simulation and inverse parameter estimation in complex porous media.