philipphahn_ms_thesis_presentation.pdf
TRANSCRIPT
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
1/44
Philipp Hahn
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
2/44
Acknowledgements
Prof. Dan Negrut
Prof. Darryl Thelen
Prof. Michael Zinn
SBEL Colleagues:
Hammad Mazar, Toby Heyn, Manoj Kumar
2
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
3/44
Outline
Motivation
Lumped Mass Model
Model properties
Simulation results
Smoothed Particle Hydrodynamics (SPH)
SPH formulation for compressible fluids Serial and parallel Implementation
Numerical Experiments
Conclusion
3
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
4/44
Technical Background and Advantages of Meshless Methods
4
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
5/44
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
6/44
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
7/44
An Intermediate Step to Meshless Acoustics
7
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
8/44
Model Properties One-dimensional model consists of
N masses which are connected via
nonlinear springs
Masses represent the inertia of
certain gas volume
Spring forces replace pressure
forces
Equation of motion for each mass:
8
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
9/44
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
10/44
Properties & Implementation In the continuum limit, linearization leads to the following differential
equation:
The model does not draw on any linearization
Second order accuracy
Implementation & visualization in Matlab
10
Example with 9 Masses and Sinusoidal Excitation
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
11/44
Pressure can be calculated from spring forces at each point
Example with 300 mass points and sinusoidal excitation:
Simulation results
Linear Springs Nonlinear Springs
11
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
12/44
Simulation results
12
Soliton Waves
Stable soliton waves can be modeled
Propagation speed of soliton waves depends on their amplitude
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
13/44
Conclusion/Limitations Pro:
Speed of sound is modeled accurately
Known nonlinear effects can be reproduced Implementation is straightforward because of the simple model
Contra:
Stability of soliton waves depends on discretization
Due to the fixed connectivity, it is not a real meshless method
The transfer in two or three dimensional implementation is challenging
Move to a more promising method, called Smoothed particle
Hydrodynamics (SPH)
13
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
14/44
For Acoustic Simulations
14
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
15/44
SPH - The Basic Idea Mainly used in hydrodynamics and astronomy
Lagrangian particle method
Each particle carries field variables (density, internal energy, velocity,) A kernel-function approach defines the influence area of each particle
15
Field variables and their
derivatives can be approximated
with the following integrations:
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
16/44
SPH - The Basic Idea With the product rule of differentiation and the divergence theorem, field
function derivatives can also be expressed by:
The surface integral is zero if the kernel doesnt intersect domain boundaries
The Integration can be approximated by a summation
16
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
17/44
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
18/44
The Achilles heel of SPH (due to the kernel approximation)
Requirements on boundary formulations in acoustics: No boundary penetration
Accurate sound wave reflection
Accurate sound excitation (moving boundaries)
No disturbances
Boundary Formulations
18
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
19/44
Dynamic Boundary Particles
19
Pro:
Easy to implement
Contra:
Moving boundaries causedisturbances
Boundary penetration is
possible
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
20/44
Mirror Particles
20
Pro:
Theoretical exact boundary
treatment due to symmetry
Less disturbances Contra:
Boundary penetration is
possible:
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
21/44
Repulsive Forces
21
Pro:
No boundary penetration
Easy to implement
Contra: Large disturbances
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
22/44
Recall
Standard SPH neglects the surface integral in the formulation for field
function derivatives
This is the root of boundary problems
Newly developed Boundary
Treatment
22
The surface integral can be used efficiently if two assumptions are made:
The field function is constant on the boundary
The field function value is equal to or slightly higher than the particle
field function value (self interaction)
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
23/44
Newly developed Boundary
Treatment
23
Pro:
No boundary penetration
Computationally efficient
Contra:
disturbances
If boundaries are smooth, a generic
solution of the surface integral can be
used in simulations
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
24/44
SPH - Implementation
24
Structure and functionality
of a basic SPH
implementations:
Leap frog integration
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
25/44
2D Implementation in Matlab
25
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
26/44
2D Hydrodynamic Tests
26
SPH liquid particles are poured with a constant initial velocity
Boundary treatment through the new developed method
Water flow into a basin
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
27/44
2D Hydrodynamic Tests
27
A square of water is discretized by 7,225 SPH particles
Boundary treatment through dynamic boundary particles
The classic SPH test simulation: Dam break experiment
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
28/44
3D Implementation on the GPU Implementation in Matlab is only reasonable for one and two-
dimensional problems
10,000 particles (One time step 30s)
Solution: Leveraging the multiprocessor architecture of GPUs
The SPH algorithm is highly parallelizable
Nearest neighbor search in parallel (radixsort algorithm)
Derivatives (interaction parallel)
Integration (particle parallel)
Three dimensional Implementation using C++ and CUDA
CUDA API (programming tools for NVIDIA GPUs)
Simulations with up to 3.5 million particles are currently possible
Speed up of about 4,000 compared to Matlab
28
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
29/44
3D Hydrodynamic Tests
29
Parameters Number of particles:
250,000
Fluid properties:Water
Boundary formulation:Surface integrals
Time stepping:t=5.0e-5s
Length:T=1.0s
Computation time:2 hours
Droplet
simulation
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
30/44
3D Hydrodynamic Tests
30
Parameters Number of particles:
82,000
Fluid properties:Water/Jelly
Boundary formulation:Surface integrals
Time stepping:t=1.0e-5s
Length:T=1.5s
Computation time:4 hours
Different
Viscosities
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
31/44
Fluid-Structure Interaction
31
Boundary formulations typically require the following information:
Boundary position
Surface normal Can be describe analytically for simple shapes
Using spherical decomposition, arbitrary shaped boundaries can be
discretized by boundary particles.
position
Surface
normal
CAD geometry
Rigid bodymotion can be
determined using
fluid-structure
forces
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
32/44
Fluid-Structure Interaction
32
Parameters Number of particles:
250,000
Fluid properties:Water
Boundary formulation:Surface integrals
Time stepping:t=0.5e-5s
Length:T=4.5s
Computation time:10 hours
Water flow on
a trough
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
33/44
SPH in Acoustic Simulations Recall the advantages of SPH:
Whole flow process is solved; (speed of sound automatically adapts to
changing fluid properties)
Complex boundaries are possible; (spherical decomposition)
No linearizations ; (based on conservation laws)
Large deformations are unproblematic; (meshless Lagrangian method)
Highly parallelizable
What needs to be analyzed:
Sound propagation
Sound excitation and reflection
Scaling and accuracy
33
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
34/44
2D Sound propagationFDTD SPH
34
SPH models sound propagation accurately Speed of sound differs slightly
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
35/44
Effects of the Smoothing Length 1D Parameter study with different smoothing lengths
Velocity excitation from the left
Analytic solution is a traveling step function Level of the step can be related to the speed of sound
35
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
36/44
3D Sound Excitation/Reflection
36
3D sound excitation in a tube
Pressure wave excitation through the moving piston
Analytic solution is a traveling step function
Modeled with 270,000 SPH particles
Boundaries are first modeled with dynamic boundary particles and then
with a combination of mirror and dynamic boundary particles
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
37/44
3D Sound Excitation/Reflection
37
Boundaries:
Pressure loss at
the edges:
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
38/44
3D Sound Excitation/Reflection
Dynamic
Boundar
y
Particles
Mirror &Dynamic
Boundar
y
Particles
38
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
39/44
Computational Efficiency
39
3D experiment with concentric
sound propagation
six different resolutions are analyzed
The pressure at 1,000 positions after
a certain simulation time is compared
with results from FDTD
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
40/44
Contributions, Limitations, Future Work
40
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
41/44
Summary of contributions Analysis of meshless Lagrangian methods with focus on applicability in
acoustical engineering
Lumped mass model of one-dimensional nonlinear sound propagation
Implementation of SPH on the CPU using Matlab and on the GPU using
CUDA
Method to model fluid structure interaction through spherical
decomposition
Analysis of the impact of smoothing length on wave speed
Analysis of sound excitation due to moving boundaries
New, surface integral based, boundary formulation
Work-precision diagram for an acoustic SPH simulation
41
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
42/44
Limitations / Future Work Limitations
Deficient boundary formulations corrupt simulation results
Particle placement is critical and results are sensitive on particle disorder
Parameters have to be chosen correctly (experience is necessary)
Future Work
Improve boundary problematic (CSPH)
Nonlinear effects need to be analyzed
Fluid-structure interaction for shock waves Hydrodynamic simulations with fluid-structure interaction
42
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
43/44
Conclusion It is generally possible to use SPH in acoustic simulations
The scaling of the GPU implementation is good
Boundary formulations need to be improved
Simulating acoustic problems is not straightforward in SPH
Exact and noise free boundary enforcement
Particle placement
Right choice of parameters
Potential applications of SPH in Acoustics: aero-acoustic problems
complex and changing domain topologies
domains with multiple propagation media
domains with high temperature or density gradients
nonlinear acoustics and shock waves with fluid-structure interaction 43
-
8/14/2019 PhilippHahn_MS_thesis_presentation.pdf
44/44