Differentiable turbulence: Closure as a partial differential equation constrained optimization

Improved turbulence closure models for large eddy simulations (LES) have the potential to impact a large variety of societal applications. This work introduces differentiable turbulence, where deep learning is embedded within a differentiable LES solver to enhance closure models given sparse observations of the true flow state. By leveraging physics-informed neural network architectures and solver-in-the-loop optimization, we put forth a technique that allows for the learning of novel closures without the use of high-fidelity numerical simulations - opening a pathway to the development and identification of LES closures in a multifidelity setting.