Deep Learning Based Speed Up of Fluid Dynamics Solvers

Download

index.pdf

Date

2022-9-8

Author

Acar, Deniz Alper

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

363
views

314
downloads

In this thesis, two distinct deep learning-based methods for the speed-up of fluid dynamics solvers are proposed. The first method called Parametric Encoded Physics informed neural network (PEPINN), is utilized to solve transient fluid dynamics problems. PEPINN is an alternative to the Physics informed neural networks (PINN) and is based on the parametric encoding of the problem domain. In PEPINN the automatic differentiation for calculation of the problem residual is replaced with finite difference kernels which improve PEPINN's time and memory complexity. This model can achieve up to 40 times speed up in wall time for the solution of the Taylor-Green Vortex problem compared to the best alternative vanilla PINN model with no loss in solutions mean squared error. It is also shown that PEPINN can be trained on up to 183 times larger data compared to the alternative vanilla PINN methods in a GTX 1080 Ti GPU. The second proposed method in this thesis is based on the hypothesis that providing the predicted solution of the steady-state Navier-Stokes equations as their initial condition might speed up the solution process. In this method, an Unet-based architecture is trained on a discretized representation of the whole problem domain given its initial and boundary conditions. The trained model is used to predict the converged solution of similar cases and the obtained results are transferred to the computational mesh of that problem. This method is tested on the steady, incompressible, subsonic flow around arbitrary airfoils.

Subject Keywords

Partial Differential Equations, Physics informed neural networks, Deep learning, Computational Fluid Dynamics, Parametric Encoding

URI

https://hdl.handle.net/11511/99620

Collections

Graduate School of Natural and Applied Sciences, Thesis

Suggestions

OpenMETU
Core

Mathematical Modeling of Turbulent Flows of Newtonian Fluids in a Concentric Annulus with Pipe Rotation SORGUN, MEHMET; Aydın, İsmail; ÖZBAYOĞLU, Evren; SCHUBERT, J J (2012-03-01) In this study, a mathematical model is proposed to predict flow characteristics of Newtonian fluids inside a concentric horizontal annulus. A numerical solution, including pipe rotation, is developed for calculating frictional head losses in concentric annuli for turbulent flow. Navier-Stokes equations are numerically solved using the finite differences technique to obtain the velocity field. Experiments with water are performed in a concentric annulus with and without pipe rotation. Average fluid velocitie...
Artificial-neural-network prediction of hexagonal lattice parameters for non-stoichiometric apatites Kockan, Umit; Ozturk, Fahrettin; Evis, Zafer (2014-01-01) In this study, hexagonal lattice parameters (a and c) and unit-cell volumes of non-stoichiometric apatites of M-10(TO4)(6)X-2 are predicted from their ionic radii with artificial neural networks. A multilayer-perceptron network is used for training. The results indicate that the Bayesian regularization method with four neurons in the hidden layer with a tansig activation function and one neuron in the output layer with a purelin function gives the best results. It is found that the errors for the predicted ...
DRBEM applications in fluid dynamics problems and DQM solutions of hyperbolic equations Pekmen, Bengisen; Tezer Sezgin, Münevver; Department of Scientific Computing (2014) In this thesis, problems of fluid dynamics defined by the two-dimensional convection-diffusion type partial differential equations (PDEs) are solved using the dual reciprocity boundary element method (DRBEM). The terms other than the Laplacian are treated as inhomogeneous terms in the DRBEM application. Once the both sides are multiplied by the fundamental solution of Laplace equation, and then integrated over the domain, all the domain integrals are transformed to boundary integrals using the Green's ident...
Numerical simulation of scour at the rear side of a coastal revetment Şentürk, Barış Ufuk; Guler, Hasan Gokhan; Baykal, Cüneyt (2023-05-01) This paper presents the results of a numerical modeling study on the scouring of unprotected rear side material of a rubble mound coastal revetment due to the overtopping of solitary-like waves utilizing a coupled hydro-morphodynamic computational fluid dynamics (CFD) model. Three cases having various wave heights are tested with six different turbulence models together with different wall functions. The hydrodynamic results (free-surface elevations, overtopping volumes, and jet thicknesses) and morphologic...
Parallel Computation of 3-D Viscous Flows on Hybrid Grids Ilgaz, Murat; Tuncer, İsmail Hakkı (2009-10-12) In this study, a newly developed parallel finite-volume solver for 3-D viscous flows on hybrid grids is presented. The boundary layers in wall bounded viscous flows are discretized with hexahedral cells for improved accuracy and efficiency, while the rest of the domain is discretized by tetrahedral and pyramidal cells. The computations are performed in parallel in a computer cluster. The parallel solution algorithm with hybrid grids is based on domain decomposition which is obtained using the graph partitio...

Citation Formats

D. A. Acar, “Deep Learning Based Speed Up of Fluid Dynamics Solvers,” M.S. - Master of Science, Middle East Technical University, 2022.