Convolutional Surrogate for 3D Discrete Fracture-Matrix Tensor Upscaling

This paper presents a 3D convolutional neural network surrogate model that accurately and efficiently predicts equivalent hydraulic conductivity tensors for fractured crystalline media, achieving over 100x computational speedup compared to traditional discrete fracture-matrix simulations while maintaining high accuracy for use in multilevel Monte Carlo frameworks.

The Gist

Imagine you are trying to predict how water flows through a giant block of Swiss cheese. But this isn't just any cheese; it's a 3D block of rock riddled with cracks, fissures, and hidden tunnels (fractures) that are invisible to the naked eye.

This is the challenge scientists face when studying groundwater in fractured rock. To get an accurate answer, they usually have to build a super-detailed computer model that simulates every single tiny crack. It's like trying to count every grain of sand on a beach to predict how a wave will crash. It's incredibly accurate, but it takes so much computer power that it's often impossible to run the simulation more than once or twice.

The Problem:
Scientists need to run these simulations thousands of times to understand risks (like storing nuclear waste deep underground). But the "super-detailed" models are too slow and expensive. They need a shortcut.

The Solution: The "Smart Guessing Machine"
The authors of this paper built a Deep Learning Surrogate. Think of this as a highly trained "AI chef."

1. The Training Phase: First, the scientists fed the AI thousands of "recipes." These recipes were the results of the slow, super-detailed computer simulations. The AI looked at the input (the rock's texture and crack patterns) and memorized the output (how fast water would flow).
2. The Magic: Once trained, the AI doesn't need to do the hard math anymore. It just looks at a new rock pattern and instantly "guesses" the answer based on what it learned.

How It Works (The Analogy):
Imagine you are looking at a complex 3D puzzle made of millions of tiny Lego bricks (the rock) and some special red bricks (the cracks).

• The Old Way (Numerical Homogenization): To figure out how water moves, you have to physically build the whole puzzle, run water through it, measure the flow, and then take it apart. You have to do this for every single puzzle variation. It takes hours.
• The New Way (The Surrogate): You show the AI a picture of the puzzle. The AI has seen millions of similar puzzles before. It instantly says, "Oh, I know this pattern! Based on my training, the water will flow at this specific speed and direction." It does this in a fraction of a second.

What Did They Find?

• Speed: The AI is 100 times faster than the traditional method. If the old method took 100 hours, the AI does it in 1 hour.
• Accuracy: Even though it's a "guess," it's a very smart one. The error is tiny (less than 22% in the worst cases, and often much lower).
• Versatility: They tested it on rocks with different crack patterns, different crack sizes, and different rock textures. The AI handled them all well.

Why Does This Matter?
This isn't just about saving time; it's about safety.

• Nuclear Waste: If we want to bury radioactive waste deep underground, we need to be 100% sure the water won't carry the radiation to the surface in 10,000 years. To be sure, we need to run thousands of simulations to account for every possible scenario. The old way was too slow to do this thoroughly. The new AI way makes it possible to run those thousands of simulations quickly and safely.
• Oil and Water: It also helps in finding oil or managing water resources in complex underground rock formations.

In a Nutshell:
The researchers created a "crystal ball" for geologists. Instead of spending weeks calculating how water moves through cracked rock, they can now use a trained AI to get a highly accurate answer in seconds. This allows them to make better, safer decisions about our planet's underground resources without waiting for supercomputers to finish their work.

Technical Summary

1. Problem Statement

Modeling groundwater flow in 3D fractured crystalline media is computationally prohibitive when using direct numerical simulations (DNS) with fine-scale Discrete Fracture-Matrix (DFM) models, especially for applications requiring repeated evaluations like Multilevel Monte Carlo (MLMC) methods for uncertainty quantification.

• The Challenge: MLMC requires a hierarchy of models with varying resolutions. Coarse models cannot explicitly resolve small-scale fractures, necessitating numerical homogenization to upscale the hydraulic conductivity of sub-resolution fractures into an equivalent tensor ($K_{eq}$).
• The Bottleneck: Conventional numerical homogenization involves solving complex Darcy flow equations on fine meshes for every block in a coarse grid, which is too slow for large-scale stochastic simulations.
• Goal: Develop a fast, accurate Deep Learning (DL) surrogate model to predict the equivalent hydraulic conductivity tensor ($K_{eq}$) directly from voxelized 3D domains representing fracture networks and matrix properties, bypassing the need for repeated numerical solvers.

2. Methodology

A. Data Generation and Physics

The authors generated a large dataset of 3D DFM simulations using the Flow123d simulator.

• Discrete Fracture Network (DFN): Fractures were modeled as square shapes with sizes following a power-law distribution ($f_s \sim Cr^{-\alpha}$). Orientation followed a Fisher distribution. Fracture density was controlled by $P_{30}$ (fractures per unit volume).
• Matrix Properties: The rock matrix was represented as a Spatially Correlated Random Field (SRF) of hydraulic conductivity tensors, generated using Gaussian fields with specific correlation lengths ($\lambda$).
• Coupling: The DFN and matrix were coupled via Robin-like boundary conditions to solve for steady-state flow.
• Homogenization: To generate training labels, the authors solved the "Anisotropy Problem" (applying pressure gradients in $x, y, z$ directions) on small cubic blocks ($15 \times 15 \times 15$) to compute the exact $K_{eq}$ tensor via numerical homogenization.

B. Dataset Construction

Three distinct datasets were created based on the fracture-to-matrix conductivity ratio ($K_f/K_m$):

1. Dataset A: $K_f/K_m = 10^3$
2. Dataset B: $K_f/K_m = 10^5$
3. Dataset C: $K_f/K_m = 10^7$

• Volume: Each dataset contains 75,000 samples.
• Input ($X$): A voxelized 3D grid ($64 \times 64 \times 64$) representing the upper triangle of the hydraulic conductivity tensor (6 channels: $k_{xx}, k_{yy}, k_{zz}, k_{yz}, k_{xz}, k_{xy}$). Fractures are projected onto this grid, and local conductivities are weighted by intersection volume.
• Output ($Y$): The 6 components of the equivalent hydraulic conductivity tensor ($K_{eq}$).
• Preprocessing: Inputs and outputs were normalized by the average matrix conductivity and standardized (log-transform for diagonal components, standard scaling for off-diagonal).

C. Surrogate Architecture

The proposed surrogate is a hybrid 3D Convolutional Neural Network (CNN) combined with Feed-Forward Neural Networks (FNN).

• Feature Extraction: Four 3D Convolutional layers ($Conv3D$) with $3 \times 3 \times 3$ kernels, followed by max-pooling and batch normalization. This reduces the spatial dimensions from $64^3$ to $4^3$.
• Regression Head: The flattened feature map passes through three fully connected layers (2048, 2048, 1024 neurons) with ReLU activation.
• Output: A final linear layer with 6 neurons (identity activation) predicting the $K_{eq}$ components.
• Training: Trained for 125 epochs using the Adam optimizer and Mean Squared Error (MSE) loss. Separate models were trained for each $K_f/K_m$ ratio due to the significant differences in data distribution.

3. Key Contributions

1. Extension to 3D: Successfully extended previous 2D surrogate modeling work to realistic 3D fractured media, addressing increased geometric complexity and higher dimensionality.
2. Stochastic Parameterization: Unlike previous works assuming constant conductivities, this model handles stochastic parameters (random fracture orientations, sizes, and spatially varying matrix conductivity), making it compatible with MLMC frameworks.
3. Tensor Prediction: The model predicts the full tensor-valued equivalent conductivity, not just scalar permeability, capturing anisotropy induced by fracture networks.
4. Scalability: Demonstrated that deep learning can replace expensive numerical homogenization in multiscale workflows without sacrificing macro-scale accuracy.

4. Results

A. Prediction Accuracy

• Overall Performance: The surrogates achieved high accuracy across a wide range of test scenarios, with a Normalized Root Mean Squared Error (NRMSE) < 0.22 for all tensor components.
• Dataset Specifics:
  • Surrogate A ($10^3$): Highest accuracy on diagonal components (NRMSE < 0.04) but struggled slightly with off-diagonal components due to broader distributions.
  • Surrogate B & C ($10^5, 10^7$): Showed more uniform accuracy across components.
• Generalization:
  • Correlation Length ($\lambda$): Models trained on $\lambda \in \{0, 10, 25\}$ generalized well to intermediate values. Accuracy dropped for very large $\lambda$ (unseen in training).
  • Fracture Density ($P_{30}$): Models performed best near training densities. Accuracy declined for extremely high fracture densities ($P_{30} > 0.002$), though remained acceptable (NRMSE < 0.22).
  • DFN Configurations: Tested on four distinct DFN configurations (including real-world data from Forsmark and Laxemar). The models maintained robust performance, indicating good generalization to different fracture patterns.

B. Computational Efficiency

• Speedup: The surrogate approach achieved a speedup of >100× compared to standard numerical homogenization when inference was performed on a GPU (NVIDIA Tesla T4).
• Cost Breakdown: While numerical homogenization is dominated by mesh generation and simulation, the surrogate cost is dominated by voxelization; the actual neural network inference time is negligible.

C. Macro-Scale Application

The authors tested the surrogates in two macro-scale problems:

1. Constraint Problem (Flow through a domain): The outflow predictions using surrogate-upscaled tensors matched the reference (numerical homogenization) with NRMSE < 0.01 for large domains.
2. Anisotropy Problem (Tensor calculation): The equivalent tensors predicted by surrogates were highly accurate (NRMSE < 0.01 for large domains).

• Conclusion: Even in scenarios where the upscaled conductivity significantly influences the result, the surrogate-based approach preserved the physical accuracy of the macro-scale simulation.

5. Significance

This work provides a scalable, high-fidelity tool for upscaling fractured porous media, which is a critical bottleneck in hydrogeological modeling.

• Enabling Uncertainty Quantification: By reducing the computational cost of homogenization by two orders of magnitude, this method makes Multilevel Monte Carlo (MLMC) simulations feasible for 3D DFM models, allowing for rigorous uncertainty quantification in radioactive waste repository safety assessments.
• Practical Applicability: The study proves that deep learning surrogates can capture complex non-linear relationships between fracture networks and equivalent conductivity, even with stochastic inputs, without requiring explicit mesh generation for every evaluation.
• Future Impact: The framework is designed to be integrated into larger multiscale simulation workflows, bridging the gap between micro-scale fracture physics and macro-scale reservoir management.