Augmenting a pure and hybrid vertical equilibrium scheme via data-driven surrogate modelling

This paper proposes a data-driven surrogate modeling approach to accelerate hybrid vertical equilibrium simulations for porous media flow, effectively reducing computational overhead while preserving physical properties like mass conservation and maintaining negligible errors compared to traditional methods.

The Gist

Imagine you are trying to predict how a giant bubble of gas (like methane or hydrogen) will move underground through a sponge-like rock formation. This is crucial for storing carbon dioxide or energy, but doing the math to track every single drop of water and every molecule of gas in 3D space is like trying to count every grain of sand on a beach while running a marathon. It's accurate, but it takes forever and requires supercomputers.

This paper presents a clever shortcut: a "Smart Hybrid" system that uses a little bit of Artificial Intelligence (AI) to speed things up without losing accuracy.

Here is the story of how they did it, broken down into simple concepts and analogies.

1. The Problem: The "Slow & Steady" vs. The "Fast & Flawed"

The researchers had two main tools to choose from:

• The Full-Dimensional Model (The "Super-Scanner"): This looks at the rock in 3D, tracking every tiny detail. It's incredibly accurate but so slow that it might take weeks to simulate a few years of gas movement.
• The Vertical Equilibrium Model (The "Flat Map"): This tool assumes the gas and water separate perfectly by gravity (gas on top, water on bottom) and flattens the problem into 2D. It's super fast, like looking at a map instead of the terrain. But, it breaks down if the gas gets stuck behind a rock or if the layers aren't perfectly flat.

The Hybrid Idea: Why not use the "Super-Scanner" only where the gas is doing something crazy (like hitting a rock), and use the "Flat Map" everywhere else? This is called a Hybrid Model.

The Catch: Even though this sounds smart, the "handshake" between the two models (where they swap data) is so computationally expensive that the whole thing ends up being slower than just using the slow "Super-Scanner" alone! It's like having a Ferrari engine but getting stuck in traffic at every intersection.

2. The Solution: The "AI Co-Pilot"

The authors realized that the "traffic jams" were caused by the computer doing the same complex math over and over again. They decided to replace these repetitive math tasks with Data-Driven Surrogates.

Think of a Surrogate Model as a crystal ball or a trained assistant. Instead of calculating the answer from scratch every time (which takes time), the computer asks the assistant: "Hey, I've seen this situation before. What's the answer?" The assistant gives a very good guess instantly.

They trained these assistants using Machine Learning (specifically simple linear regression and spline interpolation) on data generated by the slow, accurate model.

3. Three Specific Upgrades

The team identified three specific "traffic jams" and replaced them with AI assistants:

A. The "Gas Bubble Height" (Gas Plume Distance)

• The Old Way: To figure out how high the gas bubble rises, the computer had to solve a tricky, non-linear equation (like solving a complex riddle) millions of times.
• The AI Fix: They trained an AI to predict the starting guess for that riddle. Instead of starting the riddle from zero, the AI says, "I bet the answer is around here." This cut the time needed to solve the riddle in half.
• Analogy: Imagine trying to find a hidden treasure. Instead of digging randomly, the AI gives you a map with a big "X" right where you should start digging.

B. The "Flow Resistance" (Coarse-Level Mobilities)

• The Old Way: Calculating how easily gas flows through the rock layers required summing up millions of tiny calculations. It was the biggest time-sink.
• The AI Fix: They built a Spline Interpolation model. Think of this as a smooth, pre-drawn curve that connects all the possible answers. Instead of calculating the curve every time, the computer just looks up the value on the smooth line.
• Analogy: Instead of measuring the height of every single step on a staircase to know how high you are, you just look at a smooth ramp that represents the stairs.

C. The "Handshake" (Coupling Fluxes)

• The Old Way: When the "Super-Scanner" and "Flat Map" swapped data, they had to calculate the density and stickiness (viscosity) of the fluids. For water, this involves complex physics formulas that take a long time to compute.
• The AI Fix: They realized that in their specific conditions, water density and stickiness change in a very simple, straight-line way. They replaced the complex physics formulas with a simple Linear Regression (a straight line equation).
• Analogy: Instead of weighing a bag of apples to know how heavy it is, you just count the apples and multiply by a fixed number. It's almost as accurate but takes a split second.

4. The Results: Speeding Up Without Breaking Physics

The team tested this on three different scenarios:

1. Pure Flat Map: Just the fast model.
2. Static Hybrid: A mix with a fixed boundary.
3. Adaptive Hybrid: A smart mix that moves the boundary as the gas moves.

The Outcome:

• Speed: The pure model got 75% faster. The hybrid models got 18% to 44% faster.
• Accuracy: The results were almost identical to the slow, accurate model. The errors were so tiny they were negligible.
• Physics: Crucially, they made sure the AI didn't break the laws of physics. The system still conserved mass (no gas disappeared or appeared out of nowhere).

The Big Picture

This paper is like upgrading a car engine. They didn't just make the car go faster by removing the brakes (which would be dangerous); they replaced the heavy, slow gears with lightweight, AI-assisted gears.

By using simple, fast AI models to handle the boring, repetitive math, they made a complex simulation run faster than the traditional "Super-Scanner" while keeping the accuracy of the "Flat Map." This means scientists can now simulate underground gas storage much faster, helping us design better energy solutions and safer carbon storage sites.

In short: They taught the computer to take shortcuts that are so smart, they don't feel like shortcuts at all.

Technical Summary

1. Problem Statement

The simulation of fluid flow in porous media, particularly for applications like Carbon Capture and Storage (CCS) and hydrogen storage, traditionally relies on solving full-dimensional (FD) mass and momentum balance equations. While accurate, these models are computationally expensive, often prohibitive for long-term simulations or sensitivity analyses.

To address this, Vertical Equilibrium (VE) models have been developed. These models assume instantaneous phase segregation in the vertical direction, reducing the problem's spatial dimensionality by one (from 3D to 2D). However, VE models are only valid in regions dominated by buoyancy forces. To handle complex scenarios involving dynamic vertical effects, hybrid schemes have been proposed, coupling FD models in critical regions with VE models elsewhere.

The Core Challenge:
While hybrid schemes aim to balance accuracy and efficiency, the coupling process between the FD and VE domains introduces significant computational overhead. In many cases, the cost of calculating coupling fluxes and reconstructing fine-level variables negates the speed benefits of the VE model, sometimes making the hybrid simulation slower than a pure FD simulation. The authors aim to eliminate this bottleneck using data-driven surrogate models without compromising physical constraints like mass conservation.

2. Methodology

The authors propose a framework where specific, computationally expensive algorithmic blocks within the VE and hybrid schemes are replaced or augmented by fast, data-driven surrogates.

A. Mathematical Framework

• Full-Dimensional (FD) Model: Solves standard two-phase, compressible, immiscible Darcy equations in 3D.
• Vertical Equilibrium (VE) Model: Integrates FD equations over the vertical dimension. It requires reconstructing the vertical profile (fine level) from the 2D solution (coarse level) to compute fluxes and mobilities. Key components include:
  • Gas Plume Distance ($z_p$): The vertical interface separating wetting and non-wetting phases, found by solving a non-linear mass balance equation.
  • Coarse-level Mobilities: Obtained by vertically integrating fine-level mobilities.
  • Coupling Fluxes: Exchange terms between FD and VE domains requiring density and viscosity evaluations.

B. Data-Driven Surrogate Strategy

The authors selected Linear Regression and Spline Interpolation based on their extremely low inference times (crucial for frequent function calls) compared to complex operator-learning methods (e.g., DeepONet, FNO).

1. Surrogate for Gas Plume Distance ($z_p$):

   • Goal: Improve the initial guess for the Newton-Raphson solver to reduce iteration counts.
   • Features: Wetting-phase saturation ($S_w$) and the Brooks-Corey parameter ($\lambda$).
   • Method: Linear regression.
   • Rationale: Replacing the entire calculation is avoided to preserve mass conservation; only the initial guess is predicted.

2. Surrogate for Coarse-level Mobilities ($\Lambda$):

   • Goal: Replace the expensive vertical integration of fine-level mobilities.
   • Features: Wetting-phase saturation ($S_w$) and wetting-phase density ($\rho_w$, acting as a proxy for pressure).
   • Method: Bicubic Spline Interpolation.
   • Rationale: Linear regression failed to maintain the necessary monotonicity and continuity required for numerical stability. Spline interpolation provides smooth, continuous transitions that preserve these physical properties.

3. Surrogate for Secondary Variables (Density and Viscosity) in Coupling:

   • Goal: Accelerate the evaluation of phase densities and viscosities, which are called millions of times during coupling flux calculations (specifically for the wetting phase using IAPWS relations).
   • Features: Wetting-phase pressure ($p_w$).
   • Method: Linear regression.
   • Rationale: Within the relevant pressure range, density and viscosity exhibit linear or constant trends, making linear regression highly accurate and extremely fast.

3. Key Contributions

• Targeted Algorithmic Optimization: Instead of replacing entire PDE solvers, the authors identify and accelerate specific "hot spots" (gas plume distance, mobility integration, and secondary variable evaluation) within established physical models.
• Preservation of Physical Properties: The approach strictly maintains mass conservation. By using surrogates for initial guesses or smoothing filters (splines) rather than replacing the governing equations, the method avoids unphysical predictions.
• Hybrid Scheme Acceleration: The study demonstrates that data-driven surrogates can reduce the computational overhead of coupling interfaces, making hybrid simulations faster than traditional full-dimensional simulations.
• Robustness to Domain Geometry: The surrogates are shown to be invariant to domain height and adaptable to different initial/boundary conditions without retraining, provided the fluid properties and rock parameters remain consistent.

4. Results

The authors validated the approach using three test cases involving methane injection into a brine-saturated aquifer:

1. Pure VE Injection: No impermeable lenses.
2. Hybrid Static Coupling: One impermeable lens; FD/VE interface is fixed.
3. Hybrid Adaptive Coupling: Two lenses; FD/VE interface adapts dynamically based on error criteria.

Performance Metrics:

• Pure VE Model: Achieved a ~75.4% reduction in total runtime. The surrogate for coarse-level mobilities alone reduced mobility computation time by >99%.
• Hybrid Static Model: Achieved a ~44.4% reduction in runtime.
• Hybrid Adaptive Model: Achieved an ~18% reduction in runtime. (The lower gain is attributed to the large proportion of the domain remaining in the computationally expensive FD subdomain).
• Comparison with FD: The optimized hybrid model (1894s) became faster than the pure FD simulation (1976s) for the adaptive case, while maintaining physical fidelity.
• Accuracy:
  • Relative $L_2$ errors in solution fields (saturation, pressure) were negligible (often $< 10^{-3}$).
  • Mass balance errors remained within machine precision ($10^{-13}$ to $10^{-16}$).
  • The surrogate-enhanced models required fewer time steps due to the "smoothing effect" of the splines, which improved solver convergence.

5. Significance and Conclusion

This work demonstrates that data-driven methods can effectively augment traditional physics-based simulators without sacrificing accuracy or physical constraints. By focusing on inference speed and targeting specific algorithmic bottlenecks, the authors successfully:

1. Revitalized Hybrid Modeling: Made hybrid FD-VE schemes computationally viable and superior to pure FD approaches in terms of speed.
2. Validated Simple Surrogates: Proved that simple models (Linear Regression, Splines) are often sufficient and superior to complex deep learning architectures for specific, low-dimensional sub-problems in reservoir simulation.
3. Provided a Blueprint: Established a workflow for identifying computational bottlenecks and selecting appropriate surrogates that balance speed, accuracy, and physical consistency (monotonicity, mass conservation).

The study concludes that this approach offers a competitive alternative to traditional full-dimensional modeling, particularly for long-term storage simulations and sensitivity analyses where computational efficiency is paramount. Future work may involve extending these surrogates to fine-level reconstruction steps or using them to predict adaptivity criteria.