Abstract
Flow Image Super-Resolution (FISR) is essential for recovering high-resolution turbulent velocity fields from low-resolution data. Traditional super-resolution methods often fail in this domain because they treat fluid velocity components as independent color channels, ignoring the underlying physical principles of fluid dynamics. This paper proposes a vision-informed architecture that incorporates Quaternion Spatial Modeling (QSM) to represent orthogonal velocity components and Dynamic Flow Convolution (DFC) to capture fluid-specific morphological properties like viscosity and surface tension. Experimental results on Direct Numerical Simulation (DNS) datasets demonstrate that the proposed method significantly outperforms existing state-of-the-art models, achieving up to a 5.35 dB improvement in PSNR for multi-velocity tasks.
Introduction
Super-resolution (SR) techniques have seen massive success in natural image processing. However, applying these models to fluid flow data—Flow Image Super-Resolution (FISR)—reveals a significant domain gap. In fluid dynamics, the "channels" of an image represent orthogonal velocity components (U, V, W) governed by the Navier-Stokes equations, not independent color intensities. Furthermore, fluids exhibit unique morphological structures driven by physical properties like viscosity. Standard convolutional kernels, designed for rigid or natural textures, struggle to reconstruct these chaotic, droplet-like turbulent structures accurately.
Methodology
The proposed "vision-informed" framework bridges this gap by encoding physical properties directly into the network architecture.
- Quaternion Spatial Modeling (QSM): Instead of standard 2D convolutions on RGB channels, the model maps U, V, and W velocities to the imaginary axes ($i, j, k$) of a quaternion field. This allows the network to learn the inherent spatial rotations and orthogonal dependencies between velocity components.
- Dynamic Flow Convolution (DFC): To model fluid morphology, the authors developed a specialized deformable convolution. DFC uses a chain decision procedure to determine sampling offsets, mimicking the effects of viscosity, and applies constraints based on surface tension to maintain physical plausibility.
- Hybrid Architecture: The overall pipeline uses a series of Flow Feature extraction Blocks (FFB). These blocks combine Swin Transformer layers for global context with DFC for local morphology and QSM for spatial relationships.
Experimental Setup
The model was evaluated using a dataset derived from Direct Numerical Simulation (DNS) of incompressible isotropic turbulence ($Re_\lambda \approx 250$). The dataset includes high-resolution $512 \times 512$ images and low-resolution versions at $2\times, 4\times,$ and $8\times$ downsampling. Two experimental settings were used: Single-velocity (29,491 training images) and Multi-velocities (9,750 training images).
Results and Discussion
The proposed method consistently achieves state-of-the-art performance across all metrics, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). Notably, in the $4\times$ multi-velocity task, the model achieved a PSNR of 44.82 dB, representing a 5.35 dB improvement over the previous best model (DAT).
Visual comparisons show that the proposed method recovers much sharper and more accurate turbulent structures (vortices and filaments) that other models either blur out or distort. Ablation studies highlighted that while both QSM and DFC contribute to performance, the DFC module provides the most significant gains, underscoring the importance of capturing local turbulent morphology.
Conclusion
By integrating physical principles—orthogonality, viscosity, and surface tension—into the neural network's mathematical operations, this work sets a new benchmark for Flow Image Super-Resolution. While currently focused on isotropic turbulence, this "vision-informed" approach could be extended to other physical domains where imaging principles and morphology are well-defined but currently ignored by general-purpose AI models.
Summary updated