Flow Activity-Based Sampling (FABaS) of 3D LES-Generated Flow Fields for AI-Based Flow Compression and Prediction

Abstract

Flow image-based deep learning methods are emerging as a promising alternative to advance the wind-resistant design of civil structures. A key milestone for their effective and realistic implementation due to ever increasing simulation sizes is flow compression. The sampling strategy is a fundamental step in any vision-based compression technique, enabling the capture of all relevant flow features that permit the reproduction of aerodynamic phenomena, including flow-induced forces and flow characteristics in the near and far wake. While reproducing the forces can be achieved by harnessing signed distance functions (SDF), the accurate reproduction of the far field requires a more refined strategy. This investigation proposes Flow Activity-based Sampling (FABaS), a strategy that combines SDFs with flow field activity data to create a combined probability distribution that automatically selects optimal samples in both the near and far fields. The methodology is tested in a 3D LES-generated flow field around a bluff body.

Omar A. Mures

Instructor

My research interests include Deep Learning, Computer Vision and Computer Graphics.