Andreas Freund

A method I created to decrease the computational cost of performing certain fluid simulations. When performing these simulations on a coarser computational grid, additional terms must be modeled to account for the loss in resolution. This is a well-studied problem in the turbulence of a single fluid but is less understood when simulating droplets. I developed a machine-learning model of these additional terms that appear at the droplet surface using neural networks in Keras. I then augmented an existing Fortran flow solver to allow it to use this model by linking it to the C++ TensorFlow libraries. This research became the second half of my PhD thesis and was published in a high-impact journal.

A method I created to better analyze how kinetic energy is stored at and transferred between different physical scales in turbulence-containing droplets. This involved processing and analyzing hundreds of gigabytes of data from numerical simulations. This method uncovered novel physical phenomena that were used to improve simulations of such flows in a later project. This research became the first half of my PhD thesis and was published in a high-impact journal.