CBET-EPSRC: Deep Learning Closure Models for Large-Eddy Simulation of Unsteady Aerodynamics

Computational simulations increasingly enable the design of lighter, more efficient, and higher-performance flight vehicles. Current computational capabilities have successfully aided many advances in aerospace design, but challenges remain in the selection of the models used to represent turbulence. Due to practical limits on computing resources, computational simulations for engineering design typically neglect the intricate features of turbulence. The models used to approximate the missing physics contain parameters that must be calibrated to data, which is challenging for unknown flows, and often have simple mathematical forms that limit their accuracy. Recently, efficient numerical methods to calibrate the parameters of complex models during flow simulations have been developed using techniques from machine learning and constrained optimization. These methods have been successful for simple turbulent flows but have not been applied to the complex flows encountered in aerodynamics. The principal objective of this project is to develop methods by which to calibrate turbulence models for simulations of practical aerodynamic flows, which will enhance their predictive accuracy for challenging configurations. The optimization methods to be developed will be broadly applicable across engineering fields, not limited to aerodynamics, and will be made publicly available in an open-source, high-performance software package.

This project will address the need for accurate, efficient computational fluid dynamics models by developing deep learning closures and optimization methods for large-eddy simulations of turbulent separated and recirculating flows. The models will be optimized over the compressible Navier-Stokes equations using an adjoint-based approach, which will enable efficient data assimilation by avoiding the need to construct high-dimensional gradients. The resulting models will enable significant accuracy improvements compared to state-of-the-art models for comparable cost, or equivalently, significantly reduced computational cost for comparable accuracy. High-fidelity numerical datasets for several wake geometries and separated airfoil flows will be generated as target data for the optimization procedure. Additionally, a new class of online optimization methods will be developed to enable dynamic, data-free closure models that will learn directly from the governing equations, and a hybrid, multiscale deep learning formulation will be developed to model near-wall turbulent flows. The scientific community more broadly is interested in leveraging large datasets and machine learning techniques; this project therefore has potential to develop methods to be widely adopted across disciplines. The resulting algorithms, methods, datasets, and codes will be disseminated to foster adoption within the aerodynamics community and across scientific disciplines.