Reusability report: Towards inflow generators for turbulence simulations through diffusion models

Abstract The use of machine learning for scientific applications (SciML) provides an appealing alternative for solving computationally-intensive tasks in physical simulations. The chaotic and high-dimensional nature of turbulent flows makes the application of these techniques to fluid-dynamics problems challenging. Nonetheless, Li et al. [1] have shown how diffusion models can be successfully trained to generate new Lagrangian trajectories in turbulent flows, correctly reproducing most statistical benchmarks across different scales. In this work, we show how the same neural-network model and training method can be adapted to generate fluctuation velocity fields from a turbulent open-channel simulation. The resulting fields are qualitatively and statistically accurate, hinting at the possibility to use this method to accelerate direct numerical simulations, for instance through inflow generators