SimCenter | Computational Fluid Dynamics, Simulation & Computational Tools, June 12, 2017

Welcome. Today is Monday, June 12, 2017 and this is the Natural Hazards Engineering 101 Webinar Series. It is intended to
provide a common knowledge base for the NHERI community for each of the primary
natural hazards in the NHERI Program. Webinars in this series, provide an
introduction to fundamental concepts and an overview of experimental and
simulation based research. Webinars will also provide an introduction to
numerical methods and computational tools used in the Natural Hazards
Engineering Research. For more information, visit the NHERI website at, where you can find links to the SimCenter and the NHERI
Learning Center. Today’s webinar is coordinated by the Natural Hazards
Engineering Research Infrastructures Simulation and Computational Modeling
Center. This webinar is supported by the National Science Foundation under Awards
1612843 and 1520817. Any statements in this webinar are those of the presenter and do not necessarily
represent the views of the National Science Foundation. Today’s presentation
is by Professor Ahsan Kareem from the University of Notre Dame.
Professor Kareem is the Robert M. Moran Professor of Engineering and the
Director of the Natural Hazards or NatHaz Modeling Laboratory at the
University of Notre Dame. He is also Co-Principal Investigator
of the NHERI SimCenter. His work focuses on characterization and
formulation of dynamic load effects due to wind, waves and earthquakes on tall
buildings, long span bridges, offshore structures and other structures via
fundamental, experimental, laboratory and full-scale measurements,
utilizing cyber and cyber-physical infrastructure. We are pleased to have Professor Kareem. providing the Overview of Computational Fluid Dynamics in Wind Engineering and I invite Professor Kareem to begin. Good morning and good afternoon
depending on which part of the country you are from. I show that some of you may
have listened to the two earlier webinars we had. The first one dealt with the simple
introduction to the wind field characteristics, and the second one was
the wind tunnel modeling of wind effects on structures, and this one will give you
an overview of computational fluid dynamics, simulation and computational
tools in wind engineering. All of us, if you recall that our first introduction to
wind was in our physics class in high school or somewhere where we saw the
Tacoma Narrows Bridge undergoing these twisting motions and eventually
collapsing. So here are some shots of similar things done digitally using
computational fluid dynamics. This is something dealing with the past and we
are trying to use modern tools, trying to reproduce what happened during that
unfortunate day. At the same time, if we look at the recent report from NSF
about human technology frontiers shaping the future, as you can see, everything is
digital. So we’ve got to start looking at the whole computational and digital
methods of finding wind loads on structures in order to take maximum
advantages of all the tools which are available with that. So this
particular talk will give some background and then basically I’ll be
dealing with computational fluid dynamics, NHERI SimCenter and Designsafe Center a Guided Tour, a CFD Case Studies, CFD Uncertainty Quantification, a little bit about
Virtual Wind Tunnel and something, if time permits, on Stochastic
Simulations. If you look at the historical perspective starting from Da Vinci.
Leonardo Da Vinci talked about turbulence and the issues with the big
eddies and the small eddies and then more recently a British meteorologist Richardson said that “big whorls have little whorls that feed on their velocity and little whorls
have lesser whorls and so on to viscosity.” So this is the dilemma
we face when we’re dealing with turbulent flows. That it’s a multi-scale phenomena
as we’ll see in this particular slide. Challenges and goals. We
have general circulation around the globe. Then we’ve got secondary circulation
and local winds, like tornadoes or hurricanes, downbursts. All these are local
events. But if you look in the, how the energy of these fluctuations is
distributed. This is the famous Van der Hoven Spectrum so most of you will
be familiar with this one and it is a low frequency with an average wind, and
this is the part normally which you call turbulence. This turbulence interacts
with all of these structures. We have bridge systems, we have tall buildings, we have low-rise
buildings, transmission line towers photovoltaic cells, wind turbines and
offshore platforms. And they create forces of quite complex nature, but
mathematically they are currently intractable. So in order to attack these
problems, NSF recently, a lot of you are aware, that NHERI Initiative has
a NHERI Center for Computational Modeling and Simulation of the Effects of Natural
Hazards on the Built Environment, and it is these are the Co-PIs and it’s being
hosted by four universities and many other senior personnel from many other
universities participated in that one. At the same time, while we produce software, that
software has to be delivered to people and the data analysis. There’s another
center, Cyber Infrastructure Center and Designsafe-ci at the University of Texas.
Again, these are the PIs involved with that and A Cyber Infrastructure for Data
Sharing and Analysis in Natural Hazards. So these are the, here’s the mechanism,
SimCenter for developing software and Design Center, besides other
assignments they have, they will be providing the tools we develop to the
community. So let’s look at, because you know people are always skeptical about
simulation, and about the experimental versus simulations. So the comparison. CFD gives an
insight to the flow patterns that are difficult, expensive or impossible to study
using experimental techniques. In the case of experiments, those of you have done them, you
know that it depends on the number of the data points you can collect, how many
sensors you have. At the same time, if you put a sensor in the flow, you disturb
the flow. So it all depends on the type of sensors you have, how many you have. You can
obtain the data in your domain whichever you are testing. At the same time,
CFD provides a very high resolution in both space and time, and it can be used
for virtually any problem and the realistic operating conditions. Of course,
there are always errors in both sides, and one of the errors is the sensors
themselves have frequency response functions that may not be able to pick up
some of the fluctuations which are happening, and at the same time the probes
could disturb the flow, as I said earlier. In case of simulation, we’ve got even bigger
issues. The modeling, discretization of our continuum space into a discretized
space, and then implementation of various numerical techniques which we will be
using. So CFD has challenges and perception also. You know, people mostly
the naysayers, say “Oh my God! Oh it doesn’t work. It is also not reliable.” There is a
reason behind that. I’m going to compare two areas Aeronautics and Aerospace and
Structural Engineering and see why this area has been so successful and why we
are still in the developing status on this side. CFD has a long history of
applications and it’s very mature in aeronautical engineering or aerospace. In
the case of structural wind engineering, it’s a short history of CFD applications
and still an evolving field. In the case of aeronautics, users are well versed in
fluid mechanics and numerical methods, while, in our case, limited background of
fluid mechanics and numerical methods. Users developed codes for specific
applications, in the case of structural engineering, we rely on commercial
software which is very general and with a broader application, and may not be that accurate
for one particular application we are trying to use. But these guys have a
bigger advantage. They have a very restricted type of flows and very simple.
We are stuck with a difficult hand here, because we deal with a broader range of
flows with a high degree of complexity, and turbulence on top of that makes life
very complicated. If down the road, we have to go from experiments to
computations, we have to make sure that the computations can provide accurate
results. Without that, it will be a difficult transition. Summarizing again,
advances in CFD with application to wind and its effects, are critical in enhancing
the performance of the built environment. Currently, we have wind tunnels and also
vortical flow facilities which can provide measures to assess the
performance of structures on the wind, but they also have some issues with
scale effects. You know, you have to use a model scale structure and when
whenever you are dealing with a nonlinear system these scalings can become difficult. Yet,
CFD offers, in principle, a most effective means of overcoming some of the
challenges and when combined with advanced data analytics approaches, has the promise
to provide data with great details that are not possible with limited
accessibility of sensors in experiments. But again, as I said, we have major
challenges because we have massively separated flows around structures
compounded by multi-scale fluctuations, and that’s why we have now the SimCenter
and Designsafe working hand-in-hand in order to come up, build a
large community of users so that we can share new advances in this particular
area. Just a little of 101 on what happens, what is multi-scale? If you
look at the fluctuations of wind, this is the spectral density function. As you can
see, bigger eddies and they keep on merging become small until they become so small
that they are just taken away, eaten away by viscosity and they disappear.
But all these scales interact with each other, and in the dynamics of the flow in
a nonlinear fashion which we will see in a minute. Here I want to show you a plan
view of a building and the wind is coming from the left and these green blobs
are the turbulence, but as you can see, due to flow separation and the wave
development, we have three regions not developed yet. This green one,
we have fluctuations in the red one, and the fluctuation to the brown one, and we have
different spectra for each. So we started with this one, but we had additional
spectra resulting from the fluid structure interaction which makes our
life complicated. If I go into 3D, here is a building and these ghosts
which are coming, these are like big eddies. In 3D, they wrap around the building, they
tilt, they twist, they stretch. They are all mathematically describable in some way,
but for simpler types of flows. And as you can see, there’s a big mass which causes
loads on the building. So in order to capture that, we have difficulty. So let’s
go to some simple case which is Aerofoil and as you can see, the flow stays
separated attached to it throughout which makes the life of the aerospace guys
quite easy. Let’s go to a little bit more
complicated body which is a simple cylander. As you can see, we’ve got big
[volumes] behind them and we didn’t see those in that case. So the problem
becomes more difficult to handle. And here comes structural engineer [to chart these bodies] and we have not only issues there, but we have issues on the side
of this as well. As all of you know, the flows are determined by
Navier Stokes Equations, and as you can see, here is a “u” and here’s a “u”. Two nonlinear [?] nonlinear interactions and that is why things become more
complicated in solving these types of problems. Just to give you a little flavor, after
seeing an equation, I think, it’s good to see some [?] here. Here is the — you’ll hear a lot about LES and RANS today. Here is a flow around buildings.
Here it is without any inflow turbulence. In this case, you have turbulence in the flow
and that changes everything in this particular case.
For those of you doing fundamental research, here is a long rectangular body and
a stationary body. In our research, sometimes you have to use force oscillations to
see what happens with the flow around the structure and the third option is
you have an aeroelastic structure which is more like freely oscillating and as you
can see what happens to this one. And as you can see, with this type of motion the
flow rate has become a lot more complicated. In this particular one, it is a
very friendly vortex [escarpment], vortex [shedding] behind a circular cylinder.
So let’s talk about what is CFD. It is an interdisciplinary area where you have to
know fields, numerical methods, fluid mechanics, IT some computer issues and
turbulence. So with these three things, we can understand and try to model things
computationally, flows around structures. If you look at this
particular diagram here, here are so many eddies, and if I want to resolve all of
them, I need to have a very fine grid. And when you have a very fine grid, that means
you need tremendous amount of computer resources. So based on what we want to
capture, there are methods developed which are, they’re able to capture and
I’ll begin quickly and then we’ll go back to the details of each. One is DNS
which is Direct Solution of the [modern] Navier Stokes-Equations. So the grid size,
in this case, is so small that we can only do some very basic problems and
that even takes a very, very long time. Then we have a large Eddy Simulation
which is all our hopes are on this particular method, and that is we move
into here and we only capture these blue ones and leave those other ones
unresolved, and now [they] use a mathematical model to try to
capture those. And the third one, which is most widely used, is RANS. It’s Reynolds
Averaged Navier Stokes Equation which averages out all the fluctuations and
makes life a little bit easier. But for structural engineering applications, it
has other limited application because it cannot capture already excessively
separated flow conditions. Let’s talk about that this DNS. Just to give you an
idea, the number of cells you need is this power to Reynolds number. So with these three
typical problems like the Space Shuttle you need 5.6 quadrillion cells
which are currently not possible for this type of problem to be solved
computationally. Large Eddy Simulation as I’ve already given you some introduction.
Here actual flow, we try to filter out these small eddies and only try to
solve these bigger eddies which are bigger than the size of my grid. And the
ones which are smaller than the size of my grid, we call them subgrid-scale and
we use some sort of a model. And there are many models. I’ve only listed a few
here which can be used. So there’s some part of the flow is captured. They are
resolved — that’s the technical word, and the [sum] part is modeled. We don’t
analyze it computationally. Here is another description of the same thing.
Bigger eddies and smaller eddies. Physically speaking, here are the big
eddies — it becomes smaller and smaller and smaller, and those of you who like
spectral descriptions of these, here is the spectra and this is the low
frequency part which we like to computationally solve or resolve it, and
this is the one we leave for modelling it. The last one is the Reynolds-Averaged
[…]. You to take care of all the fluctuations. But once you do Reynolds
Averaging of the Navier Stokes Equations, new terms results which are
Reynolds stresses and it is the modeling of the Reynolds stresses which is the
primary function of RANS model and I’ve listed also a few of those here and
there are many more available in the literature. Just to see, from a spectral
description point of view, if the fluctuations[ in my] flow are decided
by this spectra, I do not resolve anything. You see, there is nothing here.
Everything is modeled. Then we have LES. The same thing. This is
what I need to model. You cut it off and you resolve this part. And this part, you
model it. So that is the basic concept behind these techniques. So let’s put
everything together on one diagram, and here we have a flow around a body. You
could either use Large Eddy Simulation, but you could use RANS here. So the
problem here is you [have] sharp edges separated flow features. You can see all
these eddies which are forming. Sorry. And if I look at slides of that here, because
the wind coming, we’ve got a separated zone, reattachment, wake. So all these things
are creating a complex situation for us. So we need to have this thing to be solved
computationally. We need to divide this whole domain into a grid. Then we have to
also worry about near the wall what happens, and there are, you know, you can have
a complete course on this particular area of study. At the same time, we have this
complete domain. We have inflow conditions. Something has to go
out also, and there are lateral boundary conditions. You don’t want to make them
too close to the body. This is our body. Just like in the wind tunnel, you don’t
want to have, constrain, the flow by keeping [Godric on the blockage] effect. Just to come back to these [cartoons]
again, if you look at [Airfoil?/Aerofoil?], depending on the angle of attack, its bluffness
increases and, you can see for a larger angle of attack, it also becomes like our
body, and we see there is a separation here. So on this, I am showing you the
bluffness going on this direction with the angle of attack in this case, and the
vertical axis tells you the model fidelity — how good the model has to be.
Ideally one would like DNS, but I think LES for this type of flow. In the middle
can we live with RANS, and in this case we can do laminar solutions. Something
more closer to us, is our circular cylinder, and here we can see with
the increase of Reynolds number, the flow features around this change, and
they make life — make this particular body more complicated, especially when we are
in the supercritical range, you have to use LES. This type of modeling would not
be very fruitful. Just again, just to give you an idea of [all thought or that] this
is another one URANS which I didn’t mention. This is a [finer] version of
RANS in which some fluctuation at low, very low frequencies are captured.
But this is LES, this is LES with the [In…and flow] conditions, and once again
all these things we have discussed so many times. You resolve certain parts
of your model, some part, and in case of RANS, you model everything. So the
quality of your results very much depends on the model, which is good.
Unfortunately, what happens is, the different regions of the flow are better
to model by one particular model, and some of the regions are better
represented by another model, which makes things a little bit more complex.
Okay, that’s CFD 101 again. We have got a problem statement, a mathematical model.
In this case, we have Navier Stokes Equations. We have to generate mesh. We
have to discretize space discretization, time discretization, iterative solver,
simulation runs, post-processing, and of course, we should not forget, that’s where
the problems are. If we don’t verify and validate our results, the numbers
won’t mean much. At the same time, we have uncertainty quantification. Just
like when we do experiments, we put some bars which shows the confidence we have
in our numbers, because we run experiments several times and then we
try to determine [that path]. But in case of CFD, we have to do the same thing
and [what we call] uncertainty quantification. Uncertain parameters, their
uncertainties propagated through the analysis and we will chat a little bit about this
[in the end]. So there, there are a lot of advances in mesh generation.
People in computer science, computer and computational methods study this.
But basically, there are two types of generation of meshes: structured or unstructured. And we can use CAD tools these days and
great generators. All sorts of things can happen. We don’t have to do
it the same old way like many, many years ago. People used to that. At the same time, we
have to use space discretization. That means spatial derivatives have to be
approximated, so they can be done by finite difference method. You know
whenever we take a simple loss in numerical methods, that is the first one
method we use. Or we use finite element method. But for fluids, it’s better that
you find a volume method is used which helps us in better capturing the flow
field. Time discretization again is very important. I think you might have heard
if you’re not doing much of computational method, explicit method,
implicit scheme. You have to worry about the stability, local time-stepping is
also very important with the quality of the solution which might be what you’re
looking for. What are the commercial softwares available? I think there are many
of you who use ANSYS Fluent and ANSYS CFX or CD. There are many others but
our focus is going to be on OpenFOAM because you [could] develop all tools in
an open source code which helps us to — we have access to the core, so we can
manipulate it. In all these cases, they’re like black boxes. You can just give input
depending on the instructions you receive from there. How we divide the
computational domain. For example, if you have unlimited space, and a plane is
going by, we need to make this domain capture that domain, and how do we do
that? The best way to do it — this a good example, by looking at the flow around
the plane, we want to capture these flow lines and if you take a very truncated
version of the space, you know these red portion is not truncated, and that
would not be very good for your results. And I think these are little things
which happen which keep on adding to final results and the unfortunate
mislabeling that the CFD is not reliable. Mesh generation. Let’s assume that this is
the domain we have. We generate a mesh. Mesh is very important. Rate of convergence,
accuracy of the solution and the CPU time required for that one.
And there are of course, there are many other criterias which help us to determine
which is a good quality. As I said before, we have a structured mesh, unstructured
mesh, and if these are the bodies we want to model, these are the differences. You
can see that certain mesh cannot follow everything as smoothly as compared to
an unstructured mesh. Things move, in case of wind engineering, so the trouble is the
mesh has to move also. So things get even one notch more complicated and we have
to use dynamic mesh. Here is an example of a circular cylinder in which it moves
and so does the mesh. And here is the case of a rectangular cylinder for the
same purpose. Then we have to worry about inflow conditions and outflow conditions.
It’s a little complicated. Let’s go through this. Like in the wind tunnel, we have a
long [fit] of part of the wind tunnel where we put [surface] softness, try to
generate flow which can mimic the flows for an open country or an
urban situation or [ in so local ]. Same thing is done in fluids, that you made this —
you let fluid go through a certain region and keep on recirculating it and
rescaling them, and once it is ready for being used you let it release into the
system. That’s one approach. The other approach is, people say why don’t
you do stochastic simulation of turbulence. But the trouble is stochastic
simulation does not obey the laws of fluids because it doesn’t know fluid is
[nor that] and once you enforce something which does not follow fluid mechanics, like
[continually patient in a sector], the problem just could start from the wrong
foot. So there are specialized techniques which help us to develop inflow
conditions into the domain of analysis. Just like a wind tunnel, your boxes, you
stack them up in sequence and we can have that talk shortly. Here are are some
examples from Professor [Tasura Tamura] from Tokyo who
has done quite large simulation of like a little urban development here,
and then you can zoom on the areas you would like to see, and as
you can see, flow around this particular building in two different levels is also
obvious from there. Then comes, you know, these are only isolated buildings. If you want to have regions, then we have to use WRF, Weather Research Forecasting
model, which sometimes is also coupled with LES, or nesting we call it, and that’s
how we can model large-scale weather systems and also embed them into our local
simulation. For example, it’s a very large domain, you keep on zooming until you
get into the area where is that of interest. You start with those inflow
conditions and then you nest it with your regular LES type of solution. So
computing time for flow simulation depends on a number of reasons.
The numerical algorithms you’re using, linear algebra, stopping criteria,
discretization parameters and the quality of simulation depends on the mathematical model and
underlying assumptions. The more assumption we’ll make, we’ll be
compromising some of the quality of the results. Stability, mesh, time-step,
error indicators, stopping criteria — all these things you have to embed into your code in
order to come up with some reliable simulations. What happens when you’re
interested in fluid-structure interaction? Well, you know, the vortex
induced vibration, the buffeting and the flutter takes place. Now the body’s going
to move, in this case. So we have to use a Transient Flow Solver which, again in OpenFOAM
the open source code we have as an option, we can use either pisoFoam or
pimpleFoam, and at the same time, you have to have a Structural Dynamic Solver
which could be a Newmark-Beta-based or state space-based method. We have to have
a Dynamic Mesh. They all interact together, and we get applications which
lead to vortex shedding, vortex induced vibration galloping and other issues. Now
I won’t go into the detail of this detail. I just want to make you aware so
that you can look for when you work in this area. The coupling scheme, there are two:
either Fully Coupled which many of us don’t use it, but the Iterative Coupling
or Sequential or Partitioned one we use, and in that case, the
trouble is we use Fully coupled then both of these solids or the structural and
the fluid model have to move on the same [time] which makes life quite more
complicated. But if you do interactively — here’s an example of Fluid Structure
Interaction in OpenFOAM. It’s a Partitioned strategy, and in this
case we have Fluid domain and here is the Structural domain. You could have a
Structural program like OpenSees or some others, and here is your OpenFOAM here
and this interacts through this interface. Where, for every time step, we
have to deform the mesh and then feed it back into there. And things get
a little more complicated when we are dealing with FSI type problems.
Post-processing. You know once we are done with that, we need to calculate many
other derived quantities some integral parameters because, if you have a local
pressure fluctuation, you may like to find the level load or the base shear,
the base bending moment. We may like local zooming which is not possible
in wind tunnels, flow visualization, and we have to use statistical tools to
systematically analyze the data and of course debugging, verification, validation
is a very fundamental thing. Here is one [cartoon] which we use for wind pressure
fluctuations on a tall building. The data would be coming from CFD or it could be
coming from a wind tunnel result. The same software is going to be used which is
available now on Designsafe where we can have an OpenSees or any other
commercial package trying to solve the dynamics of this building and the
pressure fluctuations coming from either a wind tunnel NHERI EF site or could be
from CFD could be fed into that and you could be getting other, you know, various
results of various kinds — accelerations, bending moments, equivalent static loads,
and the power spectral density functions. It’s a lot of processing and this
particular software will help people to do that rather than everybody
reinventing the wheel, in this particular case. We are already going, hopefully in five years,
we are going from isolated clusters to communities, like, for example, this is one
building in Houston we tested [upper error] again.
From isolated buildings, we like to go to clusters and we like to go to cityscapes
where we can have, [see] flow around the whole city. Many of you taken undergraduate
fluid mechanics classes, we use flow visualization with smoke, and then try
to capture that smoke with the camera or our naked eye. I think down the road, and as even
now, we are having this digital flow visualization. As you can see in this
particular diagram, where all the conical vortices, the horseshoe and the conical
vortex at the top can be captured digitally by releasing, or following a
particular particle, like it is shown in this particular cartoon as well.
Verification and validation is very important. Without that, we cannot trust
these values, and I think here is an example again from our Japanese
colleagues. They have done something in the the wind tunnel and then they developed a CFD
model of that one. They had the luxury of 1 billion Mesh Cells, 6,000 more than
6000 Parallel Cores on their famous Japanese K Computer which they developed
this virtual wind tunnel, and they tried to compare those. And, as you can see, without
going into every detail here, you can see the comparison between LES and experiments are
pretty good — a dashed line and solid line. And here is LES versus the experiment for
the maximum values and this is same thing for the minimum values. So for that
reason the Japanese Building Code or Standard now allows designers to use CFD
to design their buildings, but there is a caveat there. It says it has to be done
under the supervision of an expert which is I think a fundamentally
important condition. What we have done here is that, if for those of you are not
familiar with OpenFOAM, it’s very easy to download all the instructions. It’s an
open format, everything is free. So Stampede is a large computer, one of the largest in
the United States which is at TACC at the University of Texas under the Designsafe. You, as a user, could directly go and use
OpenFOAM and use the TACC resources. And here are some of the pre-processing you
have to do. All these — system, constant and polyMesh — are the fundamentals which are
required when you are trying to use OpenFOAM. Once you have that, you have to
load your things on one of the directories which is available. Then you
log in nodes in a batch mode and then then you let the computer nodes solve that, and
and like this one, we were lucky here. We are having all the results which are
possible, and this is a great computational resource available. For
those people who are not at that level yet, to be directly going on to Stampede,
we prefer that they go through the Designsafe system where you have to —
here’s where you are, with the OpenFoam values you have on your system, then
there is a Data Depot. So you transfer all your information into
various files here, as you can see the name constant and the system is here. These files
are stored here. Then there’s a workspace, Discovery Workspace in Designsafe, where
you select various solvers and all that and the program runs and your end result
is thrown back into the Data Depot and you can then suck it back into your system
[and try to ? it]. This is a little bit more friendly and more organized, and I think
most of the beginners and average users should be using this particular form.
All you have to do is go to and you can, if you don’t have an account,
you can open up an account and then you are free and ready to go.
As you can see, OpenFOAM is up there in the system. It’s all loaded, you don’t
have to do anything up there. So at this time, you have these three
things on, we are delivering on Designsafe. This OpenFOAM through the Designsafe-ci,
or you can go directly to the Stampede, or we have also a Vortex-Winds
Module has been loaded which provides you aerodynamic loads on buildings of
various shapes. There are several databases that have been merged together behind the scene.
You don’t have to worry about it. You pick up a particular geometry,
and you will get those results. [Tokems?] is another issue. This is more
like a new developments. The question is that we have already discussed so many times
that RANS is on the low end of fidelity of the quality of results. DNS is
way on the top and in-between is LES. A famous professor, late professor Ferziger —
we benefited personally from him he was a professor at Stanford, one of the
leaders in CFD. He said, “if it turns out that LES can be
done on very coarse grids, it will be one of the few times the nature has been
kind to us with regard to turbulent flows.” So the best thing is that even if
you want to use LES, it is computationally very demanding. So
the issue here is, which I’ve been thinking, that if we develop some sort of a transfer function between the result some RANS and LES
for a class of problems, like in this particular case, the RANS is a low
fidelity simulation, it’s computationally very efficient, so that’s why we love it.
But it has a large modeling error. At the same time, LES is very computationally demanding
but has higher accuracy. So if we develop a meta-model which relates these two, we
can be able to run RANS, but get the quality of LES. I think this is going
to be some of the future where model fidelity can be taken advantage of. Here
is some example. We have utilized it already in one of our previous NSF
projects where we are looking at shape optimization of buildings. So we
start with the number of initial cases, and the trouble is you have to run
simulation for each case. We decided that we develop a meta-model or a surrogate
model which relates RANS to LES. What we did — we ran many RANS
and some, the purple or magenta ones, are the LES. Based on that, through
[Craig Ing], we could develop a model. And here is that model surface, where
we run RANS, it comes in there and the output is given is LES. So for a
certain class of problems, and one can develop such things for many classes of
problems, and eventually through machine learning they could cover the entire
domain of computational fluid dynamics. That’s down the road, but there’s a good
initial success, in this particular case. If somebody wants to do real, deep
research, for example, we did several years ago with 45 million mesh cells. You
can really look at the nitty-gritty of flow around these bodies, coherent
structures, or you can even cut slices and look at the structure of flow around
the body. But this is for those people who are really interested in the fluid
dynamics of fluid structure interaction issues. Now comes the question of
uncertainty quantification in CFD. There is, you know, most of you are familiar
with aleatory variability which is natural randomness in the process, and we
cannot control it and one of them is an inflow free stream turbulence and its
characteristics. The other one is Epistemic which one is a Model-form
uncertainty and this is the discrepancy between the model and the physical
reality due to lack of information or knowledge. So our job is that we have to
propagate these uncertainties in Inflow, in Boundary conditions, in Initial
conditions, in Model uncertainty so that they enter our numerical
simulation and are present in our solution. There are various methods. Some
are very sophisticated model Intrusive methods. They are not possible yet for turbulent
flows where [interacting with a body], but for simple types of problems, they have been used.
The trouble is your deterministic problem is recast into a stochastic
version in order to analyze and one of them is a polynomial chaos, but there are
Non-Intrusive methods, and this is the very simple method, keep on running, it’s
Monte Carlo simulation, but in case of throwing a dice or a coin it’s very easy
to do it 1000 times, but if I have to run a CFD 1000 times,
that’s very demanding and challenging. So there are other ways to overcome this
difficulty. One of them we call it as a stochastic emulation. One way is that we
develope a surrogate or some surface which helps us to determine for infinite
number of inputs very quickly we can get very fine number of results which can be
then statistically analyzed in order to get and quantify the uncertainty. Dakota
is again open source code available from Sandia Laboratories. It’s
available on the net. It has all these methods embedded and there and one can
connect and link them and and also we are linking them with OpenFOAM which
will help us to do some of these things which I’m saying here without actually
running into the programming effort of all those. Here is an example. We did the —
let’s do using RANS because that’s easier and faster to run, even that took a lot
of time. So we use turbulence intensity and length scale as our two uncertain
parameters and we wanted to see how they effect our bluff body. Again, we are doing
shape optimization with this stochastic emulation. We’ve got length scale, and [?] — two variables. And then, we have two other variables which help us
to change the shape of the building. So we go through this whole surrogate
calibration and in the end we [resolve] this surrogate model here, or response
surface, if you wanna call it. This is for the Mean solution, this for the Mean plus
one Standard, and this is for Most probable solution, and then we do the
optimization and the end result is the various Pareto fronts where these are our
optimal values. If we are interested only in Mean solution, we can stop here. If you
want to go this one plus Standard deviation that’s here, or if you want to
go for the most probable solution. So that’s how the uncertainty which began
into our system can be eventually propagated and quantified in the end in
regarding our results, and in this case, this is the left coefficient and this
the drag coefficient, and these are the optimal shapes for those particular
quantities of interest. A little bit about Virtural Wind Tunnel. Several years ago, we
started this small pilot project with developing in a prototype scale. It’s a combination of SketchUp, Facebook and OpenFOAM. It
provides — you can convert structural geometries to CFD cases. You can file share
[with] the community. If you have any trouble, you can send it to your colleague here.
[I’m going to trouble can you] look at it. Access to CFD tools. OpenFOAM is
embedded in here and analysis of results. So this is a unique tool for
establishing the Virtual Community of Users who can talk to each other and
share the results. Is the shared software, shared hardware, collective knowledge and
also crowd sourcing is becoming part of it. The various case management — you can
have different structures which you are picking up. When it’s running again you
also have some little gadgets, try to see what is happening with your simulation.
And the best thing about this one is the tutorials because you’ve got — everybody’s
not fully trained in the numerical methods and fluid mechanics, so we have
provided and we are developing still more tutorial dos and don’ts. And we
have been very successful through a crowdsourcing experiment, we found out
that people who have never heard the word CFD before, and have no knowledge of
engineering, they were able to come up for the right drag coefficient for the
Empire State Building. Then you can do a lot of post-processing.
Perot you can help you to do the flow visualizations on others. How good it is?
So we decided, you know validation is always important. The Virtual Wind Tunnel
— we went into the box, and circular cylinder, squares. These are our results,
these are some of the experimental results and you will agree with me that
these are pretty good. One of the best papers on flow around
square cylinders and its drag coefficient variations is Lee (1975)
and here is our [radiation] the blue one is the Virtual Wind Tunnel. Again, realizing
the inflow conditions being slightly different, it is still a very, very good
comparison with the results. We have done some more extensive validation and
verifications using NACA Airfoil using two models: SA – Spalart and Allmaras and SST. Those of you that have done these models are advanced models, and as you
can see, I want to show you that we did many verification studies and is
supported in this paper, but I just wanted to have you look at –this is experiment 1009100778 — is the lift coefficient with OpenFOAM,
and drag coefficient is almost the same in that particular case. So one has to
get involved and develop the culture of trying to verify some results and once
our software and the fluid schemes which we use are validated enough, then we can
go beyond and looking into more cases which we would like to analyze. So the
goal is eventually, the very [mild peer point] is that I’m sitting as a
researcher or a design engineer. I don’t need to go anywhere. I should be able to
do everything digitally. I should have damping database available
to me. I could have computational platform which we are just defining. Data-
enabled design, meteorological data. If I need to do a wind tunnel experiment, at one of the
EF sites, I could do it through tele-experimentation and one could have all
these resources coming into this particular gentleman or lady to
analyze a structural tall building or any low-rise building or any other
structure which is of interest. One should not forget about other
stochastic simulations because they also come in handy in analyzing and in modeling
some of these results we’ve obtained. For example, we have
model-based approaches to [drumming] loads. They could be random in time, frequency
or time frequency. We can use Volterra based models. They have the [second order
kernel]. Once you have these kernels, we can do a lot of things with them to find
the response of structures. Then there’s a data-driven — everything in this
world is not data-driven, AI-driven, so the same thing applies to our work also.
We can use data-driven models to present stationary, non stationary,
nonlinear systems. For example, here is a hysteresis loop. It’s not a structure
under seismic loading. This is a bridge with a different angle of attack, and then
you bring it back, it does not follow the same path because there’s a hysteretic effect,
and these things have to be modeled and one could always use machine
learning and artificial neural networks. So this — the stochastic end and the data
analytics end is very important. This is eventually coupled with our CFD things
to come up with some of the world answers to some of the problems we are
still facing. Then we have cyber-physical infrastructure. We have to do cross
platform validations. Computational [versus] EF sites. We can do Hybrid/Hardware in
the loop type of experiments in which certain things which we cannot
physically model in the wind tunnel or in the CFD domain, we can do that physically and
then provide input to it through the wind tunnel or a CFD experiment. We can do
conditional simulation based on measurements. For example, in a wind
tunnel I have about 10 [hot films] to give me velocity fluctuations
at the site. I could use that and stochastically simulate fluctuations
and wind moving the measured values at those particular locations. And at the
same time, we can drive actuation systems for shaking of a frame, or a cladding
system from either a wind tunnel or CFD [outposts]. Where we are I think, just to
summarize, for us, the SimCenter and the Designsafe portals. You have direct,
access, as I showed you, or we can use through their portal. I was looking at these
Multi-Fidelity Models which I think is going to be the future, combining
them with Machine Learning tools. OpenFOAM-Dakota linkage will happen and
we will have all these things [buried] into there. So every analysis they’ll do
will automatically also provide us uncertainty quantification.
Virtual Wind Tunnel — I talked more about the advantages of that. A CFD-based input
will go into performance based wind engineering, and of course all of these things
are being done in close collaboration with the two Centers. I think I wanted to
also say that you need to remove the fear out of CFD. Most people are afraid
of CFD and I think that should not happen, and for that reason I think the Virtual Wind Tunnel, although it’s sophisticated tool, yet it will also provide as a tool
to play on your iPad. Try to have different shapes and try to make changes
and to see what small changes you make how that affects the results. So I think
this would be a very good way of developing a large database, people
running their cases and trying to deposit them in a database, and as you
can see more recently there is a game Mozak, in which people — the crowd helps
the scientists to provide input for mapping the brain cells. It’s at
University of Washington. I think we need to have — we will like to have more digital
learning hubs, so that people are given short lectures on fluid mechanics and
the dos and don’ts of what we need in computational fluid dynamics. Of
course, there are a lot of chat rooms on CFD but most of them are not applicable to wind
type of problems because people are mostly doing aerospace research. So I think we
need to make chat rooms which are more usable for the wind engineering community,
developing collaborations, validation and benchmarking, and hybrid use of NHERI EFs
for cross-platform validation. And I think with this slide I would the end
with the vortex shedding in nature. These are the migrating whales, and as you can
see when they flip they create the vortex shedding, very similar to what
they see behind the circular cylinder. And this becomes a nice environmental
engineering problem as well. With that I would thank you for your attention and
would be happy to follow it up with questions either today or later.
My email is Kareem, my last name, and we can have a dialogue,
because we want to develop, cultivate a community of users for these tools so
that we can advance our capability of analyzing structures exposed to natural
hazards. Thank you very much. Thank you Professor Kareem. Thank you for
the presentation and all the material. At this time, we will open the question and
answer session. Attendees are reminded that questions should be submitted
through the chat panel and sent to the moderator. We will relay these questions to
Professor Kareem this way. And so we do have a couple questions. Let’s start off
with wind tunnels. What role have wind tunnels played in the development
of computational fluid dynamics? As I said, wind tunnels are an essential tool
at this stage when we are developing this [scheme] for validation of our
simulations. So even in aerospace industry, although in certain areas, they
are not mature enough that they don’t use wind tunnels, but again when they are
looking at the maneuvering of aircraft they do CFD and then they go back into
the wind tunnel to develop more confidence in the robustness of their
computational simulations. So wind tunnels are still going to be an essential
component of wind engineering research. They’ll go at this stage
hand-in-hand, but eventually my goal is, and my vision is, that just like
aerospace engineering, CFD would go ahead and wind tunnels would be at the same time
only providing some validations. But how long is that going to take, I not sure.
Okay, along those lines, what are the advances needed in physical testing
to provide sufficient validation for CFD? Yeah, I think this is
some of the drawback, I mentioned. At this time, although we can measure
pressures at five, six -hundred [taps] on a building, but we cannot do the same thing
with the flow field around the structure. So if we put any sensor around the
building, that disturbs the flow itself. So what we are measuring is not the [nascent]
flow, but it’s more a disturbed flow. So I think we need to develop techniques.
I think [Laser Doppler Anemometry] is a very helpful, and more advances in those would
come hand-in-hand, and I think the more advanced sensors we have, and their
placement in the wind tunnel, the better it would be. Okay, moving outside of wind
tunnels, are there opportunities for improving models by using field data,
collected using accelerometers, placed on buildings, anemometers and pressure
gauges, or are wind tunnels providing the information we need? I think there have
been many studies about low-rise and high-rise that there are good
comparisons between what we measure in full-scale and then there are some
questions there also. I think mostly the low-rise buildings have more issues then
we have for the high-rise buildings, but I think full-scale, there [has] always
welcome and very useful and I think we cannot discard that, it’s a critical component
of validating the wind tunnel also. So all three things going hand-in-hand would
be a very useful trio. Okay, regarding inflow turbulence in CFD simulations.
There are a couple questions. Is it not true that the smallest scales tend to
dampen quickly and are less important than large scale inflow characteristics,
and what is the threshold in inflow turbulence scales in this regard?
Oh that’s a good question. The trouble is that if we do not model small scale, I mean
that becomes easier for our simulation, but the trouble is we are
losing the physics of that. As you have seen from our experiments, that the
small-scale flow scales double [in numbers] creates some of the modifications in the
flow around the structure, and if we omit modeling them, we will be seeing results
which are somewhat compromised. So one has to make sure that you can carry on
as far as you can, but again we cannot go the DNS, Direct Numerical Simulation.
So some place you’ll have to use some sort of model to take care of the smaller
eddies in there. Okay. Small scale turbulence can be generated within the
domain by flow obstacles such as buildings, vegetation. How far from the
inflow boundary do you need to be, such that the inflow turbulence may become
less important or how true is this? Anything which is in advance of the
structure is going to influence the flow field. You know there is one boundary
layer flow which is in the building codes — a very nice clean profile. But in
the real world, that clean profile is sort of an average quantity. If you have some
structures in front of your building, they are going to modify it. So it’s better
that one does a model — anything which comes in front of it or have some
indirect effect of that intuited in there. And I think, once again, there is no hard
and fast rule. We will sort of use the same strategy we used in wind tunnels that
that many blocks of the city has to be modeled in the advance of the structure.
Okay. How small are the meshes that we’re using currently to model
buildings? It depends on what — as I see that the Japanese had used 1 billion
cells for a typical of that city block, so that gives you an idea how many, what
is the size. Again, it will depend upon the nature the problem we’re
dealing with. What is the size of the domain you have, and there are many factors
which determine these numbers and [arons?] number also affects them.
And what is the role of the virtual wind tunnel in moving CFD forward to be
used in design practice? It is a more convenient tool for those people who
don’t have all the background, because it holds your hand and also tells you
you can go this way and not this way. And behind that is a very powerful OpenFOAM
simulation engine behind it. So we can do sophisticated analysis, but right now our
capabilities are limited. As the time goes, we will be going into more advanced
possibilities of structure. Our goal is that we could have a SketchUp, those
of you that you have used it. You can just sketch anything and make a building or
surroundings and they should be used in there. Right now we are still — we are
doing 2D only but very, by the end of this year, we’ll have 3D capabilities
with designed inflow conditions will be coming into the picture. Okay,
and in regard to computational fluid dynamics, what are the most obvious
questions that need attention? And by this, the question is what is the low-hanging
fruit for NSF grant seekers? Grand secrets? People writing proposals. What
are the questions related to CFD that need answering? Well I think primarily I’d
summarized in my lectures some of the the key nature of the thing. First thing is — what model
are you going to use and if you are going to use RANS, for example, you’ve got to
justify that that is good enough for the type of problems you have, and you can
justify your statements with some results which you have found in
literature where they are justified. Or if you want to use LES, I don’t
think you need any justification, but then I think you also have to provide
some justification. What type of subgrade scale model are you going to
use with that one? And what type of uncertainty you have to worry about. Then
grid generation — if it’s a static structure, if it’s dynamic structure, what type of
dynamic mesh schemes are you going to use? So there are a lot of things. If you
want to use CFD for your research, it has to be very well researched, in
order to have a good proposal. Okay. Well, we are at the conclusion of today’s Natural
Hazards Engineering 101 webinar. On behalf of the attendees, thank you
Professor Kareem for taking the time to share “State of the Simulation:
Practice in Wind Engineering”. To the attendees, thank you for your
participation and questions. The NHERI SimCenter is shifting the focus from
wind engineering to tsunami engineering for upcoming NHE 101 webinars. Register
for these on the Designsafe web site, and check your email inbox for emails
from [email protected] that will have registration links for these.
If you aren’t receiving these email announcements, check your Designsafe
account settings to include these types of announcements. Thank you for attending
today’s webinar. Thank you. you

Leave a Reply

Your email address will not be published. Required fields are marked *