Towards a machine-learning-based large eddy simulation of offshore wind farms

Abstract

This study introduces a Scale-Adaptive Machine-Learning Subgrid-Scale model developed to predict subgrid-scale turbulence within the framework of large eddy simulations for offshore wind farms. Unlike traditional subgrid-scale models that rely on blending of isotropy and scale similarity, the proposed approach leverages a supervised learning framework based on physically informed flow observables derived from mixed modelling theory and Leonard decomposition. The model employs a novel encoder–decoder neural network architecture designed to capture coherent enstrophy dynamics and multi-scale turbulence interactions. Skip connections and latent representations serve as implicit filters, enabling the model to represent both structural and functional aspects of turbulence. Trained using data from a scale-adaptive LES method, outcome of the presented model has been validated for its ability to learn and reproduce key turbulence characteristics, such as intermittency and energy transfer, across resolutions and flow scenarios. A-priori tests confirm its capacity to capture statistical turbulence features, while a-posteriori tests demonstrate that the model dynamically predicts eddy viscosity and produces flow fields comparable to high-resolution LES with traditional SGS models. When applied on coarser meshes, the model maintains accuracy, as evidenced by agreement in the ratio of subgrid to total kinetic energy. These findings support the potential of this machine-learning-based model as a physics-aware, scalable modelling approach for complex turbulent flows.