Home // INFOCOMP 2019, The Ninth International Conference on Advanced Communications and Computation // View article

Automatic Generation of Adjoint Operators for the Lattice Boltzmann Method

Authors:
Stephan Seitz
Martin Bauer
Negar Mirshahzadeh
Andreas Maier
Harald Köstler

Keywords: Lattice Boltzmann method; Computational Fluid Dynamics; Adjoint Methods; Gradient-based Optimization; Tensorfow

Abstract:
Abstract—Gradient-based optimization techniques for computational fuid dynamics (CFD) have been an emerging feld ofresearch in the past years. With important applications in industrial product design, science and medicine, there has been an increasing interest to use the growing computational resources in order to improve realism of simulations by maximizing their coherence with measurement data or to refne simulation setups to fulfll imposed design goals. However, the derivation of the gradients with respect to certain simulation parameters can be complex and requires manual changes to the used algorithms. In the case of the popular lattice Boltzmann method (LBM), various models existsthat regard the effects of different physical quantities and control parameters. In this paper, we propose a generalized framework that enables the automatic generation of effcient CPU and GPU code for optimization using symbolic descriptions of arbitrary lattice Boltzmann methods. The required derivation of corresponding adjoint models and necessary boundary conditions are handled transparently for the user. We greatly simplify the process of CFD optimization for a broader audience by providing LBM simulations as automatic differentiable building blocks for the widely used machine learning frameworks Tensorfow and Torch.

Pages: 41 to 46

Copyright: Copyright (c) IARIA, 2019

Publication date: July 28, 2019

Published in: conference

ISSN: 2308-3484

ISBN: 978-1-61208-732-0

Location: Nice, France

Dates: from July 28, 2019 to August 2, 2019