Generative super-resolution of turbulent flows via stochastic interpolants

Area of Science:

• Fluid Dynamics
• Computational Science
• Machine Learning

Background:

• Turbulent flows present multiscale challenges due to limited experimental resolution and high-fidelity simulation costs.
• Coarse flow representations often fail to capture crucial fine-scale dynamics, limiting practical applications.
• Generative models offer a potential solution for reconstructing unresolved flow scales.

Purpose of the Study:

• To leverage generative models, specifically stochastic interpolants, for super-resolution of turbulent velocity fields.
• To reconstruct unresolved fine-scale dynamics from low-resolution conditional data.
• To evaluate a novel patch-wise application of stochastic interpolants for efficient flow reconstruction.

Main Methods:

• Employed stochastic interpolants for super-resolution of a two-dimensional turbulence case study.
• Utilized an iterative, patch-wise application of stochastic interpolants for efficient reconstruction.
• Compared the patch-wise approach against full-field methods and other generative models.

Main Results:

• The patch-wise strategy successfully reconstructed physically consistent super-resolved flow snapshots.
• Key statistical quantities, including the kinetic energy spectrum, were accurately recovered.
• Stochastic interpolants outperformed competing generative models across various metrics.

Conclusions:

• The patch-wise application of stochastic interpolants provides an efficient and effective method for turbulent flow super-resolution.
• This approach yields results comparable in quality to full-field methods.
• Stochastic interpolants show significant potential for advancing the analysis of turbulent flows.