Hyper-Parameter Analysis of Deep Auto Encoder for Flow Prediction

EasyChair Preprint 5572, version 2

Abstract

Reducing the state of the system from high order dynamical space to a low dimensional subspace has been a challenging task. Linear Projection methods have been used extensively in the past for building such Reduced Order Models. Though they have been successful in modelling various several non-linear scenarios, their usage limits their applications pertaining to High order dynamical systems effectively. In this study we aim to make use of advancement made in the field of Deep Learning to build a DL based ROM. We aim to probe the impact of hyperparameters pertaining to flow prediction using deep autoencoders built with the help of artificial neurons. We make use of different network sizes and sizes of the time-stamp as two of our parameters to compare the performance in flow prediction. Dataset used here was generated using in-compressible URANS CFD simulation for simulating Von-Karman vortex street at Reynolds' number 100 around a bi-dimensional cylinder using OpenFOAM solver icoFoam.

Keyphrases: Hyperpararmeters, deep autoencoder, von Karman Vortex Street

@booklet{EasyChair:5572,
  author    = {Prakrit Tyagi and Pranav Bahl and B B Arora},
  title     = {Hyper-Parameter Analysis of Deep Auto Encoder for Flow Prediction},
  howpublished = {EasyChair Preprint 5572},
  year      = {EasyChair, 2021}}