Abstract: In this talk, I will present a new data-driven methodology which improves upon the accuracy of state-of-the-art large eddy simulation (LES) approaches. The proposed methodology relies on learning the components of the subgrid stress (SGS) tensor in a particular flow-dependent reference frame - the strain-rate eigenframe. These components are typically smooth, thus state-of-the-art machine learning strategies are well suited for learning them. These components are also invariant under a change of coordinates, so the proposed methodology automatically yields SGS models that are Galilean and frame invariant provided the model inputs are themselves Galilean and frame invariant. The Buckhingham Pi theorem can further be employed to arrive at unit invariant SGS models with a reduced input space. Lastly, the proposed methodology yields SGS models whose stability characteristics are easily established, and in particular, it enables the construction of energy-stable SGS models. I will present numerical results illustrating the methodology’s capability to generate simple, accurate, and efficient data-driven SGS models that generalize to turbulent flows outside the training set.
Biography: Dr. Evans is the Jack Rominger Faculty Fellow and an Assistant Professor in the Ann and H.J. Smead Department of Aerospace Engineering Sciences at the University of Colorado Boulder. Dr. Evans received his Ph.D. in Computational and Applied Mathematics from the University of Texas at Austin in 2011, and he worked as a postdoctoral fellow at the Institute for Computational Engineering and Sciences between 2012 and 2013. Dr. Evans’s research focuses on the modeling and simulation of complex fluid flows and fluid-structure interaction. Additionally, Dr. Evans is actively involved in the the development of fully integrated design, analysis, and optimization technologies using the framework of isogeometric analysis.