Convolutional Surrogate for 3D Discrete Fracture-Matrix Tensor Upscaling
A new 3D CNN model predicts groundwater flow in fractured rock with over 100x speedup on GPU.
Researchers Martin Špetlík and Jan Březina have published a novel AI-driven approach to dramatically accelerate groundwater flow modeling in complex 3D fractured media. Their work, detailed in the paper "Convolutional Surrogate for 3D Discrete Fracture-Matrix Tensor Upscaling," tackles a major computational bottleneck in geoscience: simulating fluid movement through fractured crystalline rock. Traditional Discrete Fracture-Matrix (DFM) simulations are prohibitively expensive, especially when repeated evaluations are needed for uncertainty quantification or parameter studies. The team's solution is a surrogate model that predicts the equivalent hydraulic conductivity tensor from a 3D voxelized domain, bypassing the need for costly numerical homogenization at each step.
The surrogate architecture combines a 3D Convolutional Neural Network (CNN) with feed-forward layers, allowing it to capture both local spatial features of fractures and their global interactions within the rock matrix. The model was trained on data generated from high-fidelity DFM simulations, with three separate versions accounting for different fracture-to-matrix conductivity contrasts. It was rigorously tested across a wide parameter space, including various fracture network geometries and matrix-field correlation lengths, maintaining high accuracy.
In practical demonstrations, the AI surrogate was applied to two macro-scale problems: computing equivalent conductivity tensors and predicting outflow from a constrained 3D domain. The results showed that the surrogate-based upscaling preserved the accuracy of conventional methods while achieving computational speedups exceeding 100x when inference is performed on a GPU. This breakthrough integrates into a Multilevel Monte Carlo (MLMC) framework, where it efficiently upscales sub-resolution fracture effects when transitioning between simulation accuracy levels, making large-scale, probabilistic groundwater assessments finally feasible.
- Uses a 3D CNN + feed-forward network to predict hydraulic conductivity tensors from voxelized fracture data.
- Achieves normalized root-mean-square errors below 0.22, preserving accuracy of traditional numerical homogenization.
- Enables over 100x faster simulations via GPU inference, integrated into a Multilevel Monte Carlo (MLMC) framework.
Why It Matters
This makes large-scale, probabilistic groundwater resource assessment and contaminant transport modeling computationally feasible for the first time.