 |
Center for Nonlinear Studies |
This talk is on a method to reduce the computation time and memory usage of Lattice Boltzmann Fluid simulations with neural networks. The method works by compressing the state size of a simulation and learning the time dependent dynamics on this compressed representation. This allows fluid simulations to be emulated by a neural network with significantly less computation and memory. In addition to presenting this method, we will discuss both our current and future work in this area. In particular, a neural network based fluid flow library that is designed to handle large scale simulations and provide an environment to quickly test related research ideas. We will also present a neural network based method of doing design optimization of airfoils in steady state flow and its possible extension to time dependent flows.
Host: Michael Chertkov