Surrogate models for Rock-Fluid Interaction: A Grid-Size-Invariant Approach

This paper introduces a grid-size-invariant deep learning framework using UNet++ architectures as efficient surrogate models for rock-fluid interaction, demonstrating their superior performance and memory efficiency over reduced-order models in predicting fluid flow through dynamic porous media with non-static solid fields.

The Gist

Imagine you are trying to predict how a drop of ink spreads through a sponge, but this sponge is made of rock, and the ink is actually carbon dioxide (CO2) being injected deep underground to fight climate change.

To do this accurately, scientists usually run massive, super-computer simulations. Think of these simulations like a high-definition movie where every single pore in the rock and every molecule of fluid is tracked frame-by-frame. While the picture is perfect, rendering this movie takes hours or even days. If you want to test 1,000 different scenarios (like changing the rock type or injection speed), you'd be waiting years.

This paper introduces a clever shortcut: Surrogate Models. These are like "AI assistants" that can watch a few minutes of the high-definition movie, learn the rules of how the ink moves, and then instantly predict what happens next without needing the super-computer.

Here is how the researchers built these AI assistants, explained through simple analogies:

1. The Two Main Strategies: The "Translator" vs. The "Pattern Matcher"

The team built eight different AI models, but they fall into two main camps:

Camp A: The "Translator" (Reduced-Order Models)
Imagine trying to describe a complex painting to a friend over the phone. You wouldn't describe every single brushstroke; you'd summarize the main shapes and colors.

• How it works: The AI first uses a "compressor" (an Autoencoder) to shrink the massive, detailed data of the rock and fluid into a tiny, simple summary (a "latent space"). It's like turning a 4K video into a tiny sketch.
• The Prediction: A second AI (the Predictor) looks at this tiny sketch and guesses what the next sketch will look like.
• The Result: The AI then "un-compresses" the sketch back into a full picture.
• The Catch: Sometimes, if you compress too much or the summary gets a little fuzzy, the final picture loses detail. The researchers tried different ways to make the "summary" more accurate, including a technique called "Adversarial Training" (where two AIs play a game of "fake vs. real" to force the summary to be perfect).

Camp B: The "Pattern Matcher" (Grid-Size-Invariant Approach)
This is the paper's big innovation. Imagine you are learning to recognize a specific type of cloud. Usually, you might only practice on small 4x4 inch photos. If you then try to identify that cloud in a massive 10-foot mural, a normal AI might get confused because the scale changed.

• The Innovation: The researchers built an AI that is scale-invariant. It learned the rules of the fluid flow on small patches of rock (like 64x64 pixels) but can apply those same rules to a massive, unseen rock formation (256x256 pixels).
• Why it's cool: It's like teaching a child to recognize a "dog" by showing them a small toy dog, and then having them correctly identify a giant Great Dane. The AI doesn't need to memorize the whole big picture; it just needs to understand the local patterns, which saves a huge amount of memory.

2. The Secret Sauce: "Rollout Training"

When predicting the future, small mistakes add up fast. If you guess the weather for tomorrow, and then use that guess to predict the day after, your error grows.

• The Problem: Standard AI training only looks one step ahead. "What happens next?"
• The Solution (Rollout Training): The researchers taught the AI to look further ahead. During training, they made the AI predict 8 steps into the future at once and corrected its mistakes based on the whole sequence, not just the next step.
• The Analogy: It's like learning to ride a bike. A normal trainer only corrects you if you wobble right now. A "rollout" trainer corrects you based on whether you fell over 10 seconds later. This makes the AI much more stable and accurate over long periods.

3. The Architecture: UNet vs. UNet++

The researchers tested two different "brain structures" for their AI:

• UNet: A standard, reliable architecture (like a classic sedan).
• UNet++: A more complex, nested version (like a sports car with extra aerodynamics).
• The Verdict: The more complex UNet++ won. It was better at capturing the fine details of how the fluid moves through the rock, especially when combined with the "Rollout Training."

4. Why This Matters

The specific problem they solved is tricky: as CO2 flows through the rock, it actually eats the rock (dissolves it), changing the shape of the tunnels as the simulation runs. Most AI models struggle with this because the "map" keeps changing.

By using the Grid-Size-Invariant approach with Rollout Training, they created a model that:

1. Saves Memory: It can run on a standard laptop GPU instead of a massive supercomputer cluster.
2. Saves Time: It predicts 100 steps of simulation in less than a second (compared to hours for the traditional method).
3. Stays Accurate: It doesn't drift off course as quickly as older methods.

The Bottom Line

This paper is about teaching AI to be a smart, efficient apprentice. Instead of waiting for a master (the supercomputer) to solve every problem, we have built an apprentice that can look at a small piece of the puzzle, understand the rules of the game, and instantly predict the outcome for the whole board. This opens the door to testing thousands of carbon storage scenarios quickly, helping us design safer and more effective ways to store CO2 underground.

Technical Summary

1. Problem Statement

The paper addresses the computational challenges associated with modeling rock-fluid interactions, specifically in the context of Carbon Capture and Storage (CCS).

• High Computational Cost: High-fidelity numerical models (based on Computational Fluid Dynamics - CFD) require fine meshes and significant computational resources to solve partial differential equations (PDEs) governing fluid flow, transport, and chemical reactions (rock dissolution).
• Limitations for Multi-Query Tasks: The high cost restricts the use of these models for tasks requiring thousands of simulations, such as uncertainty quantification and optimization.
• Specific Challenge: The interaction involves non-static solid fields (rock dissolution changes the porosity and geometry over time), making it difficult to use the solid field as a static mask for correction.
• Memory Constraints: Training deep learning models on large, high-resolution 3D datasets is often limited by GPU memory availability.

2. Methodology

The authors propose and compare two main categories of surrogate models (machine learning approximations of the physical system) trained on a 2D dataset of CO2 injection in a carbonate reservoir (generated via GeoChemFOAM).

A. Reduced-Order Models (ROMs)

These models use a two-stage pipeline to reduce dimensionality:

1. Compression: A Convolutional Autoencoder (CAE) compresses the high-dimensional physical fields (CO2 concentration, porosity, x/y velocities) into a low-dimensional latent space.
   • Comparison: The authors compare Traditional Autoencoders (AE) vs. Adversarial Autoencoders (AAE). The AAE uses a discriminator to regularize the latent space distribution, aiming for better stability.
2. Prediction: A neural network predicts the evolution of latent variables over time.
   • Comparison: UNet vs. UNet++ (a nested architecture with skip connections).
   • Training Strategy: The predictor is trained autoregressively (predicting $t+1$ based on $t, t-1, t-2$).

B. Grid-Size-Invariant Framework

This is the paper's novel contribution. Instead of compressing the data, it uses a single fully convolutional neural network (UNet or UNet++) that possesses grid-size invariance.

• Concept: Because the network uses only local convolutional operations (no fully connected layers), it can be trained on small spatial patches (e.g., $64 \times 64$) and infer on much larger, unseen domains (e.g., $256 \times 256$) without retraining.
• Rollout Training: To prevent error accumulation in long-term autoregressive predictions, the authors employ rollout training. The model is trained to minimize loss over a sequence of $T$ timesteps simultaneously, rather than just the next single step.
• Curriculum Learning: Training starts with single-step loss and transitions to rollout loss to accelerate convergence.
• Boundary Conditions: A penalty term is added to the loss function to enforce boundary conditions.

3. Key Contributions

1. Grid-Size-Invariant Surrogate Models: Demonstrated that training on small subdomains allows for inference on larger domains, significantly reducing memory consumption during training while maintaining high accuracy.
2. Architecture Comparison: Systematically compared UNet vs. UNet++, finding that UNet++ consistently outperforms UNet in both ROM and grid-size-invariant frameworks, particularly for capturing fine-grained details in fluid flow.
3. Training Strategy Optimization: Showed that Rollout Training significantly improves long-term prediction stability and reduces error accumulation compared to standard single-step training.
4. Adversarial Training Analysis: Investigated AAEs for compression. While AAEs offer better latent space regularization, they are harder to tune and did not consistently outperform traditional AEs in the long-term prediction phase of the ROMs.
5. Application to Dynamic Systems: Successfully applied these methods to a system with dynamic geometry (rock dissolution), proving that data-driven models can handle non-static solid fields without explicit geometric masks.

4. Results

The models were evaluated on a dataset of 32 simulations (24 for training, 8 for validation) using Pearson Correlation Coefficient (PCC), Structural Similarity Index (SSIM), and Mean Squared Error (MSE).

• Performance of Grid-Size-Invariant Models:

  • The UNet++ with Rollout Training achieved the best overall performance.
  • Generalization: These models outperformed the baseline (trained on full domains) on unseen validation data, demonstrating superior generalization and reduced overfitting due to the data augmentation effect of subsampling.
  • Metrics: Achieved PCC > 0.90 for CO2 concentration and porosity, and > 0.75 for velocities over 100 timesteps. SSIM scores indicated high visual fidelity.
  • Memory Efficiency: Training on $64 \times 64$ patches reduced peak GPU memory usage by a factor of ~4–7 compared to full-domain training (e.g., 0.39 GB vs 3.44 GB for UNet++).

• Performance of ROMs:

  • ROMs with UNet++ performed similarly to the full-domain baseline for concentration and porosity but lagged in velocity prediction.
  • AAE vs. AE: AAEs provided better initial latent space regularization but did not sustain a long-term advantage over traditional AEs in the prediction phase; in some cases, traditional AEs performed better after 50 timesteps.

• Efficiency:

  • Inference Speed: The surrogate models are orders of magnitude faster than the CFD solver (GeoChemFOAM). While the CFD solver took ~3 hours per simulation (24 CPUs), the surrogate inference took < 1.5 seconds (1 GPU).
  • Training Time: Grid-size-invariant models trained faster and with less memory than full-domain models, though rollout training increased memory usage slightly during the training phase.

5. Significance and Future Work

• Scalability: The grid-size-invariant approach solves a critical bottleneck in applying AI to fluid dynamics: the inability to train on large 3D datasets due to memory limits. This framework enables the use of high-resolution data on standard hardware.
• Hybrid Simulation Potential: The authors propose a future hybrid workflow where the surrogate model runs for a number of timesteps to accelerate the simulation, and the system reverts to the high-fidelity PDE solver only when prediction metrics (e.g., PCC) drop below a threshold.
• Broad Applicability: While tested on CO2 storage, the methodology is generalizable to any time-evolving fluid flow system governed by PDEs, including 3D inertial flows and other multiphase interactions.

In conclusion, the paper establishes that grid-size-invariant neural networks trained with rollout strategies offer a robust, memory-efficient, and highly accurate alternative to traditional high-fidelity simulations for complex rock-fluid interaction problems.