Prediction of Multiscale Features Using Deep Learning-based Preconditioner-Solver Architecture for Darcy Equation in High-Contrast Media

Modeling subsurface fluid flow in porous media is crucial for applications such as oil and gas exploration. However, the inherent heterogeneity and multi-scale characteristics of these systems pose significant challenges in accurately reconstructing fluid flow behaviors. To address this issue, we proposed Fourier Preconditioner-based Hierarchical Multiscale Net (FP-HMsNet), an efficient hierarchical preconditioner-learner architecture that combines Fourier Neural Operators (FNO) with multi-scale neural networks to reconstruct multi-scale basis functions of high-dimensional subsurface fluid flow. Using a dataset comprising 102,757 training samples, 34,252 validation samples, and 34,254 test samples, we ensured the reliability and generalization capability of the model. Experimental results showed that FP-HMsNet achieved an MSE of 0.0036, an MAE of 0.0375, and an R2 of 0.9716 on the testing set, significantly outperforming existing models and demonstrating exceptional accuracy and generalization ability. Additionally, robustness tests revealed that the model maintained stability under various levels of noise interference. Ablation studies confirmed the critical contribution of the preconditioner and multi-scale pathways to the model’s performance. Compared to current models, FP-HMsNet not only achieved lower errors and higher accuracy but also demonstrated faster convergence and improved computational efficiency, establishing itself as the state-of-the-art (SOTA) approach. This model offers a novel method for efficient and accurate subsurface fluid flow modeling, with promising potential for more complex real-world applications.

💡 Research Summary

The paper introduces FP‑HMsNet, a novel deep‑learning framework designed to predict multiscale basis functions for Darcy flow in high‑contrast porous media. The authors identify the limitations of traditional multiscale numerical methods (e.g., MsFEM, GMsFEM) – namely, high computational cost, difficulty handling extreme heterogeneity, and slow convergence under large permeability contrasts – and propose a hybrid “preconditioner‑solver” architecture that merges a Fourier‑based spectral preconditioner with a hierarchical multiscale convolutional network.

The Fourier preconditioner transforms the high‑contrast permeability field κ(x) into the frequency domain, normalizes the spectrum, and feeds the result into the learning pipeline. This operation preserves global high‑frequency information while dramatically improving the conditioning of the underlying linear system. The hierarchical multiscale network then processes the preconditioned data through several levels, each mimicking the construction of local snapshot spaces and spectral reductions used in mixed GMsFEM. By learning both global Fourier features and local multiscale representations, the network captures low‑frequency flow patterns and fine‑scale heterogeneities simultaneously, overcoming the spectral bias typical of standard Fourier Neural Operators (FNO).

Theoretical contributions include a proof of stability via Lipschitz‑constrained layers, convergence guarantees showing exponentially decaying residuals, and error bounds of the form ‖error‖₂ ≤ C·h^α + C’·N^‑β, where h is the fine‑grid size and N the number of Fourier modes. The preconditioner is shown analytically to reduce the condition number of the discretized system from O(κ_max/κ_min) to O(1), cutting the number of conjugate‑gradient iterations by roughly 30 % even when κ varies by four orders of magnitude.

For data, the authors generate a large synthetic dataset using Karhunen‑Loève expansion (KLE) to create log‑normal permeability fields with prescribed correlation length and variance. They compute mixed GMsFEM multiscale basis functions on a 30 × 30 fine grid and a 10 × 10 coarse grid, resulting in 102 757 training samples, 34 252 validation samples, and 34 254 test samples. The dataset spans 5 matrix permeability values, 7 fracture permeability values, and up to 25 fractures per realization, yielding 875 distinct configurations.

Training employs the Adam optimizer with an MSE loss, L2 regularization, and learning‑rate scheduling. On the test set, FP‑HMsNet achieves an MSE of 0.0036, MAE of 0.0375, and R² of 0.9716, outperforming baseline models: standard FNO (MSE ≈ 0.0123), Hierarchical Attention Neural Operator (HANO, MSE ≈ 0.0087), and pure GMsFEM (≈12 % relative error). Robustness tests add Gaussian noise (σ up to 0.1) to inputs; the MSE increase remains below 0.0004, demonstrating strong noise tolerance.

Ablation studies reveal the critical role of each component. Removing the Fourier preconditioner raises MAE to 0.062 (≈66 % increase) and degrades R² to 0.928. Replacing the hierarchical multiscale pathway with a single‑scale CNN similarly reduces performance, confirming that both global spectral conditioning and local multiscale feature extraction are essential.

Computationally, FP‑HMsNet runs in O(N² log N) time and sub‑linear memory with respect to the domain size N², thanks to the spectral preconditioning and efficient convolutional implementations. On an RTX 3090 GPU, inference takes ~0.018 s per sample, roughly five times faster than a conventional mixed GMsFEM solve and twice as fast as the best existing neural operator models.

In conclusion, FP‑HMsNet delivers state‑of‑the‑art accuracy, speed, and robustness for high‑contrast Darcy flow problems, bridging the gap between physics‑based multiscale methods and modern operator learning. The authors suggest future extensions to three‑dimensional domains, non‑Newtonian flow, integration with real field sensor data, and deployment on low‑power edge devices for real‑time reservoir monitoring.