 |
Center for Nonlinear Studies |
Turbulence is pervasive throughout most flows of scientific and engineering interest, yet the modeling of turbulence and its effect on engineering designs remains a persistent challenge. This talk introduces a data-driven machine learning and optimization framework that enables new insights into turbulence physics, modeling, and applied engineering design. In the first part of the talk, a new autonomic closure for coarse-grained turbulence simulations is introduced and examined in a priori tests of large eddy simulations (LES). The autonomic closure is a nonparametric approach that learns a flow-specific closure on-the-fly, and is fundamentally different from previous approaches rooted in continuum mechanics. The closure is found without external training data by solving a supervised learning problem in the inertial range by using powerful kernel-based machine learning techniques. The autonomic kernel closure is remarkably accurate in a priori tests, and shows promise as a tool for data-enabled model discovery. In the second part of the talk, high-dimensional optimization is applied to the design and control of practical thermal-fluid systems. A canonical example of turbulent flow in engineering practice is found in utility-scale wind farms. We demonstrate a novel optimization of the turbine layout and control settings using powerful adjoint techniques. Adjoint simulations are used to efficiently find gradients of high-dimensional, nonlinear wind farm flow simulations. This enables a groundbreaking increase in the fidelity of the fluid flow simulations used in wind farm layout and control optimizations. The talk is concluded with a discussion of how data-driven machine learning and optimization can enable insights into new physics and aid in reduced-order model development.
Host: Misha Chertkov