Accelerating urban scale simulations leveraging local spatial 3D structure

Abstract

This paper presents a hybrid methodology for accelerating Computational Fluid Dynamics (CFD) simulations intertwining inferences from deep neural networks (DNN). The strategy leverages the local spatial data of the velocity field to leverage three-dimensional convolutional kernels within DNN. The hybrid workflow is composed of two-step cycles where CFD solvers calculations are utilized to feed predictive models, whose inferences, in turn, accelerate the simulation of the fluid evolution compared with traditional CFD. This approach has proved to reduce 30% time-to-solution in an urban scale study case, which leads to generating massive datasets at a fraction of the cost.