U-FNO: Hybrid U-Net & Fourier Neural Operator

• The paper introduces U-FNO, a hybrid neural operator that fuses a global Fourier transform for long-range interactions with a U-Net convolutional pathway for localized feature extraction.
• It demonstrates significant performance gains with approximately 46% lower gas saturation errors and 24% reduced pressure errors compared to CNN benchmarks.
• U-FNO delivers up to 10^5 times faster inference while requiring only one-third the training data, making it ideal for efficient, real-time simulations in complex geoscience applications.

The U-Net enhanced Fourier Neural Operator (U-FNO) is a hybrid neural operator architecture engineered to accurately and efficiently approximate nonlinear solution operators for complex multiscale partial differential equations, particularly in the context of multiphase flow in porous media. U-FNO integrates the global, mesh-independent modeling power of the Fourier Neural Operator (FNO) with the localized, multiscale feature extraction of a U-Net-inspired convolutional pathway. This composite design yields improved generalization, data efficiency, and fidelity in highly heterogeneous domains where capturing both nonlocal interactions and sharp localized fronts is essential.

1. Architectural Design and Operator Formulation

The U-FNO architecture extends the classical FNO by introducing a dual-path operator layer combining:

• Global path: A spectral integral operator, implemented via Fourier transforms, encapsulates long-range interactions and cross-domain correlations.
• Local path: A U-Net convolutional operator extracts spatially localized, fine-grained features, enabling the network to resolve sharp fronts and high-frequency details.

Mathematically, a single U-Fourier layer operates by

$$v_{m_{k+1}}(x) = \sigma\Bigl(\mathcal{K}v_{m_k}(x) + \mathcal{U}v_{m_k}(x) + W(v_{m_k}(x))\Bigr),$$

where:

• $$\mathcal{K}v_{m_k}(x) = \mathcal{F}^{-1}\bigl(R \cdot \mathcal{F}(v_{m_k})\bigr)(x)$$ is the global Fourier operator,
• $\mathcal{U}v_{m_k}(x)$ is the output of the U-Net convolutional branch,
• $W$ is a linear local channel-mixing transformation, and
• $\sigma$ is a nonlinear activation function.

Input fields (e.g., spatial permeability, porosity maps, scalar injection parameters) are first “lifted” by a fully connected network $P$ into a high-dimensional latent space. After $M$ iterations of intertwined global-local operator layers, the output is projected back to the physical space by a projection network $Q$.

2. Mathematical Rationale and Model Properties

The U-FNO is designed to approximate a nonlinear solution operator

$\mathcal{G}^\dag: \mathcal{A} \to \mathcal{Z},$

where $\mathcal{A}$ is the space of admissible inputs (e.g., field maps and physical parameters) and $\mathcal{Z}$ is the space of desired outputs (e.g., gas saturation or pressure fields).

The global Fourier path leverages the mesh-invariance and long-range coupling of the Fourier basis, aligning with the mathematical structure of elliptic and parabolic PDEs. However, absence of local modeling can result in excessive smoothing—a regularizing effect inherent to global spectral truncation—limiting the network's ability to learn sharp solution features. The U-Net local path compensates for this by hierarchically extracting and reconstructing localized, high-frequency solution content, similar to methods used in state-of-the-art image segmentation and dense prediction networks.

The resulting hybrid update preserves the mesh-free, generalizable qualities of FNO while substantially enhancing expressiveness and training accuracy in scenarios where solution regularity is low or physical phenomena are strongly localized.

3. Performance Benchmarks and Data Efficiency

U-FNO exhibits marked improvements over pure FNO and conventional convolutional neural networks (CNNs), as demonstrated on CO$_2$-water multiphase radial flow problems in geoscience:

• Gas saturation prediction: U-FNO achieved plume area mean absolute errors approximately 46% lower than CNN benchmarks and higher $R^2_{plume}$ values (0.981 vs. 0.955 for CNNs), indicating superior spatial accuracy in plume-front detection.
• Pressure buildup: The model reduced field mean relative error by roughly 24% compared to CNNs.
• Data efficiency: U-FNO required only about one-third as much training data as CNNs to reach comparable accuracy, an essential property for applications with expensive or limited simulation data.
• Computational speed: Once trained, U-FNO delivered end-to-end inference speeds up to $10^5$ times faster than conventional PDE solvers (e.g., ECLIPSE), facilitating real-time or large-scale uncertainty quantification.

The composite operator design also mitigates typical overfitting issues encountered with purely convolutional surrogates in high-dimensional, heterogeneous settings.

4. Real-World Applications in Multiphase Flow and Geoscience

U-FNO has been validated as a surrogate model for rapid, mesh-independent simulation of multiphase CO$_2$—water injection in subsurface storage, a domain characterized by extreme heterogeneity in permeability, porosity, anisotropy, and highly localized nonlinear phenomena:

• Inputs: High-resolution spatial maps (e.g., permeability, porosity), scalar injection parameters, and well configurations.
• Outputs: Time-resolved spatial fields of gas saturation and pressure buildup, including critical localized features such as advancing plume fronts and pressure transients near wells.
• Practical benefits: Orders-of-magnitude speedup, data-parsimonious training, and improved generalization render U-FNO suitable for computational risk analysis, assisted history matching, geostatistical inversion, and “digital twin” applications.

The architecture is notably more robust and data-efficient than CNN surrogates, which typically require extensive data and can overfit or fail to generalize in the presence of geologic complexity.

5. Extensions, Variants, and Model Evolution

Several lines of subsequent research have built upon or extended the U-FNO framework:

• IU-FNO and Variants: Introduce implicit recurrent Fourier layers for deeper, parameter-efficient operator learning, further stabilized by a shared global update and U-Net refinement. These models demonstrate even longer-term predictive stability and reduced parameter count, particularly in turbulent flow prediction (Li et al., 2023, Wang et al., 5 Mar 2024, Zhang et al., 4 Nov 2024).
• Fourier-MIONet: Adapts the U-FNO decoder in a multi-input architecture, achieving comparable accuracy with 90% fewer parameters, much lower memory requirements, and faster training by separate treatment of time and spatial coordinates (Jiang et al., 2023).
• Task-specific Hybrids: Hybrid strategies exploit domain knowledge, such as applying Fourier operators in periodic directions and U-Net in non-periodic ones (HUFNO), or combining global-local FNO branches (GL-FNO) to address domain-specific scaling and feature separation (Du et al., 21 May 2024, Wang et al., 17 Apr 2025).
• Performance in Other Physics: U-FNO and its principles have informed surrogate development for phase field dynamics, turbulent channel flows, and micromechanics, though in some scenarios (e.g., sharp hemodynamic gradients (Zou et al., 8 Apr 2025)) explicit residual-based U-Nets may outperform UFNO.

A persistent trend is that the integration of U-Net mechanisms (multiscale, skip-connected convolution) with Fourier-operator based modeling improves expressive power, especially in multiscale or data-limited regimes.

6. Methodological Implications and Future Research Directions

U-FNO substantiates the value of global-local hybrid operator learning for PDE-based systems. Its ability to combine mesh-independence, data efficiency, and flexibility in handling highly variable local features opens avenues for:

• Generalization across discretizations: U-FNO’s mesh-invariance supports adaptive gridding and transfer across different simulation resolutions.
• Multiphysics and Multiscale Operator Learning: The operator-splitting paradigm sets a foundation for extending neural operator architectures to coupled multiphysics and multi-domain problems.
• Training and Optimization: Empirical results highlight the benefits of Fourier-local hybridization for gradient flow and efficient learning, though further investigation into regularization, stability losses, and explicit physics-informed constraints is warranted.
• Deployment: U-FNO’s efficiency, especially for repeated-query and uncertainty tasks, suggests a role for integration in operational digital twins, rapid scenario screening, and probabilistic inversion frameworks.

Emerging model variants explore physics-informed losses, attention-based local enhancer modules, and hierarchical variants to further improve physical consistency and generalization in even more challenging environments.

7. Summary Table of Key Quantitative Results

Metric	CNN Benchmark	FNO	U-FNO
Gas Sat. $R^2_{plume}$ (test)	~0.955	not stated	0.981
Gas Sat. MAE (test)	reference	higher	~46% lower
Pressure Buildup MRE (test)	reference	higher	~24% lower
Training Data Needed (rel.)	$1\times$	higher	~0.33×
Inference Speedup vs Simulator	n/a	n/a	$10^5\times$

In all metrics, U-FNO improves upon both CNN and FNO baselines, with lower error, higher $R^2$, dramatically higher inference speed, and stronger data efficiency in high-heterogeneity PDE settings.

---

U-FNO’s architectural innovations and consistent performance across complex, heterogeneous, and data-limited operator-learning tasks establish it as a significant advancement over both classical FNO and conventional deep convolutional surrogates for scientific operator learning.