Project Description: Scientific evidence strongly indicates that human activities are playing a significant role in bringing about major changes in our climate system, particularly in the last five decades. Significant changes in climate as a result of these human activities are projected for the rest of the 21st Century and beyond. The scientific analyses also strongly indicate that the globally-averaged temperature charges are being largely driven by changing atmospheric concentration of radiatively important gases and particles, lending credence to concerns about much larger changes in temperature, precipitation, and for changing probability in the likelihood for severe weather events over the coming decades. Numerical models of the physics, chemistry, and biology affecting the Earth-atmosphere climate system are key tools to both understanding the past changes in climate and to making the projections for the future that are necessary for planning and policy purposes. Today, even as we prepare to run these complex models of the Earth's climate system on petascale machines, we realize that despite the extensive capabilities that petascale will enable, a number of critical limitations in modeling the climate system require an exascale capability. The purpose of this presentation is to discuss the science underlying the concerns about climate change, the models of the Earth's climate system used in studying climate change, and the challenges being faced in taking these models to petascale and exascale. The NCAR Community Earth System Model (CESM), the most cited and used climate model in the world, will be the basis for the discussion on climate modeling capabilities now and into the future. Speaker Bio: Donald J. Wuebbles is the Harry E. Preble Professor of Atmospheric Science at the University of Illinois. He is a professor in the Department of Atmospheric Sciences as well as an affiliate professor in the Departments of Civil and Environmental Engineering and in Electrical and Computer Engineering. He was the first Director of the School of Earth, Society, and Environment at Illinois, was the first Director of the Environmental Council at the University, and was Head of the Department of Atmospheric Sciences for many years. Dr. Wuebbles is an expert in numerical modeling of atmospheric physics and chemistry. He has authored over 400 scientific articles, relating mostly to atmospheric chemistry and climate issues. He has been a lead author on a number of national and international assessments related to concerns about climate change. He has also been a lead author on national and international assessments relating to atmospheric chemistry and the effects of human activities on stratospheric and tropospheric ozone. Dr. Wuebbles and colleagues received the 2005 Stratospheric Ozone Protection Award from the U.S. Environmental Protection Agency. He has been honored by being selected a Fellow of three major professional science societies, the American Association for the Advancement of Science, the American Geophysical Union, and the American Meteorological Society. He is the Chair of the Global Environmental Change Focus Group for the American Geophysical Union. He shares in the 2007 Nobel Peace Prize for his work with the international Intergovernmental Panel on Climate Change. He was a member of a federal advisory committee that assessed and in 2009 published a report on the potential impacts of climate change on the United States. Professor Wuebbles is a Coordinating Lead Author for the next major international IPCC assessment of climate change that will be published in 2013 and is a leader in the next U.S. National Climate Assessment, being a member of the Executive Secretariat and the Federal Advisory Committee.
and aerosol effects, and several representation
of ice and other things that we know need improvement.
I'll talk a little bit about those.
But another key aspect is increasing
the resolution of the model.
And that requires more computational resources.
And that's where the x-scale and the peta-scale
particularly come in, because it's
allowing us to change the resolution in ways we
couldn't even consider before.
So meeting all these objectives that we need to have,
this can't be accomplished with current systems.
That's why we need to go from peta-scale to x-scale.
This is just a representation of all those different components
of the Earth system model that we need to consider.
And then enhanced treatment of those components
are going to be necessary as we look at our enhanced
understanding we need for the future
to be able to provide policymakers
a better understanding.
We're already seeing some aspects of that
as we do make improvements in the model.
The CAM-5 is the atmospheric module in the CSM,
the latest version.
The 5 denotes that it has micro-scale representation
of the aerosol and cloud processes
that were not in the previous models.
And this graph, which I didn't realize
how unclear it was in the distance,
does do a much better job in comparison for things
like looking at El Nino events and the energy in an El Nino
event compared with previous models.
And we can see that even more clearly
that we compare the CAM-5 version of CSM
with the older models, which were public community climate
system model before they changed the name to Earth system model.
4 and 3.
And generally, when compared to the observations on the left,
the new model does a much, much better job
of representing the right timing and duration of an El Nino
event, of a grand slope events in general,
El Nino and La Nina events compared to earlier
versions of the model.
Another key aspect with the new CAM-5
is that it does a much better job
in dealing with dry zone biases, which
means those parts of the world which
tend to be extremely dry.
And in capturing those subtropical dry zones,
the CAM-5 model has shown on the lower right,
compared to the observations, does a much, much better job
than the earlier models.
And a lot of that is because of the improvements
in treating the clouds and aerosol processes
and the reactions.
It should start automatically.
We also find that as we increase the resolution of the model,
that we start capturing physics that we could not fully
capture before.
And this shows an example of looking
at precipital water for a version of the model.
This is an earlier version of the model of CAM-4
that was run at an eighth of a degree.
This is only a quarter degree on Blue Waters right now.
It was run for a fairly relatively short period
of time.
But in doing so, we found, as you
can see on the central right or on the side
here, that it can capture the production of typhoons
and hurricanes much better than you could
with a lower resolution model.
So to capture those kind of large events more accurately,
we need to go to extremely high resolution.
Well, to go to that kind of resolution,
we need to change the central core of the model,
the dynamical core, the way the model was
written under the Earth's sphere.
And we've got to transition over the years
as the resolution has changed in the CSM and its predecessors,
going particularly from finite volume, which
we ran that for quite a long time,
to the new model, such as we're running here at Blue Waters.
The home spectral element approach
that will allow us to go to about a quarter degree.
It might be able to take us to about an eighth of a degree,
but perhaps not beyond that.
And eventually, we want to be running
a one and two kilometer resolution.
And I'll talk about that shortly.
And so the newest approach is called M-PASS.
And I never remember the form.
It's triodal or something like that scheme
that's involved there.
It's the grid on the right-hand side
with black around the Earth that allows
us to be able to actually go to non-hydrostatic, which
means we're no longer making the hydrostatic assumption
in the equations of motion, so that we can go down
to an extremely high resolution and be
able to get down to those few kilometer kind of scales,
where we can actually capture cloud processes
without having to parameterize them.
So as we think about exascale and the kind of problems
we want to run in the future, we know
that the current models, running at a one degree resolution
typically, cannot capture the detail at the regional scale.
What's going on in Pacific locations in the US,
for example, that we want to have a better understanding of.
And therefore, we're going to need
to go down to a much higher resolution, quarter
of a degree or a tenth of a degree.
Eventually, a tenth of a degree or better
will probably be the standard.
We're going to get a lot of increased accuracy.
To give you just a sense of that,
I can show you a whole lot more slides of what increased
resolution does for you.
And as a result, we'll be able to get a much better
understanding of the potential impacts of climate,
down to the scale that humans and natural systems really
care about.
As part of that, we also need to run
a large number of ensembles, because we
want to account for, which means a large number of runs
for any given problem, because we
want to detect the signal relative
to the natural variability.
And we can do that just by doing a number of runs that give you
different representations of the natural variability
at any given time.
And then you can look at probably that range
to detect what real signal is coming out
of that due to the human forcing that is occurring.
So large ensembles are important,
because they help us to robustly determine
simulated natural variability.
They allow us to better assess this question of climate
extremes that I mentioned, particularly
at Pacific locations, and their changing likelihood,
and to investigate detection and attribution of climate
changes in the 20th and 21st centuries
that we can then associate with human activities
relative to those natural variations
that we see in the climate system.
As part of that, eventually, we're
going to want to go to models that can fully
resolve cloud processes.
To do that, we need to get down to a 1 to 2 kilometer
resolution, where the cloud's physics can
be represented accurately.
And we get away from the parameterizations
that we currently have to use.
And for that, we're definitely going
to require excess skill.
This shows the kind of risk, with just
looking at the physical resolution of the model,
as you go from 200 kilometer resolution,
which is the kind of models we have used,
even in the current IPCC assessment, previous models,
the previous assessment, we standardly
used the 2 degree, or 200 kilometer resolution.
We're currently running in blue water,
so the 25 kilometer resolution, and this
is showing the West Coast.
And what we want to get to eventually
is to the point where we can really
capture in great detail what is going on
with the mountains and other parts of the geography
in places like California.
So looking at excess skill, what kind of problems
would we want to run on that?
Well, I'm giving you some hints about that already.
Obviously, we're going to want to run
analyses of climate change and projections going out
into the future.
This century, and likely well beyond this century,
climate change does not end at 2100,
even though we tend to only go to graphics up to that point.
Those analyses are going to be a big fully coupled.
We include all aspects of the climate system.
We're going to run around 200, 400 years or more in length.
The inputs will be time dependent.
We'll look at a range of different future emissions
scenarios, a large number of ensembles.
And to do that, we're going to have to have machines
like Exascale to really get there.
At the same time, one of the biggest problems
in the weather climate community right now
is this time period between the weather scales,
that are roughly, and we can now predict the weather roughly
out about a week.
And we can project climate changes out
over the next 20 years or so.
But we do a very poor job in understanding
what's likely to happen in the next decade.
And yet, that's extremely important to farmers and others
that would really like to know that information.
What's the weather likely to be like a year from now,
so they can do a better analysis of how they should
plant their crops, et cetera?
Well, weather forecasting is generally
done as an initial value problem.
It's how well you know the initial conditions that
allow you to go up a week.
And eventually, the ultimate is to get out in about two weeks,
and then before chaos in the system takes over.
Climate modeling is typically a boundary value problem.
We're trying to look at long-term time scales
and look at how the climate is changing over those time
scales.
We need to understand how those changes,
and what kind of changes we want to discover
over a decade or less time scale.
And that's going to require a very different way of thinking.
That's where I think it's an exciting area of research
to look at weather climate analyses over the next decade.
That's where I would be putting my efforts.
I'm probably going to retire before that happens.
But eventually, this is really important.
And to do that, we're going to have a much better understanding
of the ocean cycles, and what roles they play,
and year-to-year weather, et cetera,
as we try to do that.
So it becomes a very difficult problem
to have the right initial data in the model
to make this happen.
There's a large number of ensembles, very high resolution
to be able to do those analyses.
Again, we're going to need exascale.
We'd also like to go back over that 2,000-year time period.
That's much longer.
It would be really interesting to go back
to 20,000-year time period to the end of the last ice age.
Or to look at some of the previous interglacial time
periods, so we have a lot less understanding of those
than we did in the last 20,000 years.
And to do that, again, it's a very long time scale analysis
for, again, requiring exascale machines.
So as we look at the needs for computing resources
over this next decade, the increased complexity
of the models as we improve the physics, the chemistry,
and the representation of biology in those models,
how we represent all the land use changes
and all kinds of things, are going
to put a requirement on that.
Going to higher resolution is going
to put a requirement on that.
How we better assimilate satellite data
to put another requirement on that.
And then the duration of the runs and the ensemble sizes
also puts another.
And so we're likely to see a significant increase
in computational resources towards really understanding
our diverse climate system and how it's likely to change.
For computer scientists, this represents critical challenges
to address scalability, programmability, resilience.
We've got the power needs of the machine itself.
That's just going to require some special thinking.
I'm going to talk about that in just a second,
because there's some interesting work at LBL related to that.
And certainly concerns about IO and memory,
putting out a huge number of output data
that we really don't know how to analyze.
Wiener et al, a couple of years ago,
and then continues to work on this at LBL,
has been looking at the question of,
should we be taking an alternative approach
to how we design exascale type machines?
Should they not just be based on having a hardware that
can generally be applied to all types of problems,
like Blue Waters is, it's a pedestal machine.
But should we actually be designing machines
that are designed for the specific problem at hand?
And so the quote from that paper,
we need to design machines to answer
specific scientific questions, rather than limit our questions
by the available machines.
So it's something to think about,
and something I think we all need to much more
thought about as we look to the future.
And that raises big questions about the power requirements
of an exascale machine, and that perhaps we
need to go this route in order to reduce the power
needs for answering specific science questions.
So I thank you, and I'll be happy to address
any questions you have.
How do you address the success of a model?
Like, how do you evaluate when the model is
considered successful?
Is this proscriptive?
Do we use older data and predict what comes to date?
Or do we predict the future and wait and see?
No, we've largely based those analysis
on looking at lots of processes first at the small scale.
You can pull out the different parameterizations or processes
in the model and look at specific questions
where you have great detailed measurements.
But then on a larger scale, the model itself,
then you can look at, well, what kind of data
do we have globally that we can look
at both in the oceans, in the atmosphere,
and in the biosphere that we can compare with.
And very, very detailed analyses are done to look at that.
But we still have an uncertainty.
If you look at what's called the climate sensitivity factor,
that's something that was developed years ago for just
doubling the amount of carbon dioxide
from pre-industrial levels, about 300 parts per million
to 600 parts per million.
That climate sensitivity has remained essentially unchanged
over the last 30 years, from about 1 and 1
half degrees Kelvin to, or centigrade,
to 4 and 1 half degrees centigrade.
And you can find different data in the past
that support either end.
And the models tend to support that part of it.
And we haven't not reduced that uncertainty
because if we gain more understanding,
we also gain more uncertainty.
And so it's just part of the process.
But we do try to do very careful analyses.
And some things the models do extremely well on.
It really does look like Earth when we look at the model.
But it's not a perfect Earth.
And we know there's some things that we don't represent well.
We know some things like clouds and precipitation patterns
models tend to do poorly.
And then they will greatly improve those
if we increase the resolution.
That's it.
So the question is, how do we use a large amount of data sets
that we have to consider?
Most of it, the model basically starts out with the solar
input and drives the atmosphere and the oceans and so forth.
The one exception to that is sub-grid scale.
So anything less than the one degree or a quarter of a degree
or whatever we're running at that we need to consider,
we will need to have a parameterization based
on a large amount of observations of the things
like clouds that have been used to develop it.
But the model basically derives everything else.
And then we extensively compare it
with the satellite observation data sets.
But we don't use them initially.
We'll have an initial condition that's
based off of a previous model run usually.
But you're not dependent on those data sets
for really running the initial model.
It's more for the comparison.
What's the size of your model?
Now, that's not true of the weather models.
They're entirely dependent on that initial condition.
What's the size of your model?
I was trying to think.
It's hundreds of thousands of lines of code.
And I don't remember exactly what the.
So we're running at a quarter degree
on blue waters with 150,000 processors.
I think we can run a year roughly every four or five
days.
It's a big animal.
I just don't remember the exact size.
One last question.
In the news, they said the European models
are better than the American models.
Certainly for weather, that has been the case lately.
And that is because of medium-range forecasts,
the European model's been doing a somewhat better job
than the American model.
A lot of that has to do with computational power.
NOAA has not received an increase
in its computation large enough compared
to the Europeans in the last four or five years
or decade.
And as a result, we've fallen a little behind
the resolution that they can run their weather models at.
In terms of the climate models, the US models
are amongst the very best.
And there's some good European ones.
There's a Japanese one that's very good.
And there's three really good ones.