Geometry Workshop Obergurgl 2021

I will present a new data-driven approach for significantly increasing the speed of yarn-level cloth animations with extremely high geometric complexity. First, we use *numerical homogenization* to learn an approximate continuum-level stress-strain relationship for woven and knitted yarn patterns. Using this continuum model, we can now effectively approximate yarn-level cloth using a FEM-based thin-shell solver, instead of simulating every individual contact between threads. Next, we propose a method to animate the individual yarns in a mechanics-aware fashion, by learning their displacements from pre-computed yarn-level deformations. These two methods combine to effectively speed up the animation of fine knitted garments by several orders of magnitude, including real-time and interactive animations of millions of threads. These ideas also create a number of new open problems in the field of cloth simulation.