Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:10
Hello, and welcome to the NVIDIA
0:12
AI Podcast. I'm your host, Noah
0:14
Kravitz. Climate change is a defining
0:17
issue of the 21st century. Can machine learning
0:19
help us model climate even though the future
0:21
will not be like the past? Can it
0:24
help us plan for coming new extremes of
0:26
heat, flood, drought, and rising sea levels? My
0:28
guest today is here to help us break it all down. Chris
0:31
Brotherton is the Senior Director of Climate
0:33
Modeling at the Allen Institute for Artificial
0:35
Intelligence, also known as AI2. And Chris
0:37
is here at GTC 2024 in San
0:41
Jose, California to talk about machine
0:43
learning and the future of climate
0:45
modeling. Chris, thanks so much for
0:47
joining the NVIDIA AI Podcast. Thanks,
0:49
Noah. So, let's start with
0:51
the basics. Before we get into your work
0:54
specifically, can you tell us a bit about
0:56
AI2? AI2 is
0:58
a group of about
1:00
200 people now who are
1:03
primarily supported philanthropically
1:05
through the Paul
1:08
Allen estate, working
1:10
on a diverse range of AI
1:13
tasks. Some of
1:15
the most important tasks include open
1:18
source, large language modeling, machine
1:20
reasoning, and knowledge
1:22
distillation. Environmental AI
1:24
is a newer focus for
1:27
AI2, but now in addition
1:29
to climate modeling, AI2
1:31
is using artificial
1:34
intelligence to detect
1:36
illegal fishing, track wildlife,
1:39
and other really interesting and
1:41
diverse AI challenges. So
1:43
we're here to talk about climate change, and everybody
1:46
is aware of it. We're in
1:48
California right now. You traveled
1:50
down from the Pacific Northwest,
1:52
so we're both pretty aware
1:54
of droughts, wildfires, record-blaking snowfall
1:57
and rainfall, just these
1:59
extraordinary things. extreme weather events seem to be
2:01
coming one after another. Can
2:03
you tell the audience what climate modeling
2:05
is, how it relates to climate change,
2:08
and how it's helping us all prepare
2:11
for whatever's to come in the future? In
2:13
the Western United States, there's always been
2:16
extreme weather. We've always had floods
2:18
and droughts. We've always had fires.
2:22
But as I think everyone who lives
2:24
here realizes, something is
2:26
changing. We're getting a
2:28
lot more smoke. It seems
2:30
like the floods are stronger. Of course,
2:33
snowpack is going down on average, but
2:35
the extremes are getting more
2:37
severe. And we
2:39
wonder what's going to happen in the future. Well,
2:42
a lot of what is going to happen
2:44
is a consequence of basic
2:46
physics, which we can encode
2:49
into mathematics. And
2:52
at the most basic level, we can
2:54
understand it on the back of a
2:56
napkin. But if we're interested in understanding
2:58
what's going to happen in places like
3:00
California, for instance, we need to
3:02
go a lot further. And
3:05
60 years ago, Suki Manabe
3:07
at a NOAA lab called the
3:09
Geophysical Fluid Dynamics Lab built
3:12
the first climate model with that in mind.
3:15
And what a climate model is, a physics-based
3:18
climate model is, is it's a
3:20
representation of the Earth, of
3:22
the atmosphere, of the ocean, of
3:25
other land surface components like land and ice
3:27
and so on, on
3:29
a grid, encoding mathematical
3:31
algorithms that represent our best
3:33
guess at how different kinds
3:35
of processes work. Atmospheric
3:38
winds, ocean currents, more
3:40
complicated things like how clouds are
3:43
formed, even how
3:45
soil decomposes vegetation
3:48
by microbes into CO2 and so on.
3:51
So there are all these things going on
3:53
in a climate modeling model, and
3:55
there's the 60 years of evolution
3:57
of complexity and using... finer
4:00
and finer grids that have made them more
4:02
sophisticated. But they are both
4:05
expensive to run. They
4:08
can only be run and developed
4:10
at major climate modeling centers. And
4:14
there's still uncertainties. They don't tell
4:16
us the whole answer. So there's
4:18
a lot further to go. And a
4:21
few years ago, I got interested in
4:23
trying to think about how machine learning
4:25
might help with that. And so climate
4:27
modeling used to be done with FORTRAN? Climate
4:29
modeling is still done with FORTRAN. OK. Even
4:31
if I said used to, I was like,
4:33
oh, no, I'm setting myself up. All right.
4:35
No. In fact, the heart
4:38
of many climate models still dates back to
4:40
the 1970s. And
4:43
in fact, it's hard to hire people
4:45
to do climate model development because they
4:47
have to learn FORTRAN. Although
4:49
the FORTRAN of today is not the FORTRAN of the
4:52
1970s. It
4:54
is a much more modern language. But
4:56
that is a hindrance. Because,
4:59
for instance, now we have computer
5:01
languages like JAX, which naturally
5:04
produce code which is friendly to
5:06
using with machine learning. But
5:08
no operationally
5:11
used climate model is written without friendliness
5:13
in mind. Right. And so that means
5:15
that the field is going to
5:18
have to be transformed in a somewhat
5:20
revolutionary as opposed to evolutionary
5:22
way because of
5:25
this legacy technology that it's based on. And
5:27
so you said a couple of years ago
5:29
you started thinking about how to use machine
5:32
learning in climate modeling. And so
5:34
what's that been like? What have you been
5:36
able to do or started thinking about doing?
5:39
And you said it's going to
5:41
be more of a revolutionary than
5:44
evolutionary process. Where both you specifically
5:46
and the field in general, where
5:49
are things at with bringing machine
5:51
learning into climate modeling? Yeah, well,
5:53
let's start five years ago
5:55
in, well, actually a little over five years
5:57
ago, I think in 2007. when
6:01
I was actually standing in front
6:03
of Paul Allen and proposing a
6:06
project on using machine learning
6:08
for climate modeling to
6:11
him and a committee of
6:13
experts in both climate
6:15
and actually in computational
6:17
science. At that time, there was an
6:19
extreme degree of skepticism that machine learning
6:22
could be helpful at all in the
6:24
near future. In fact,
6:26
so much so that our project
6:28
was originally rejected and only later
6:31
got revitalized after Paul
6:33
Allen's passing and turned into a
6:35
small venture with initially a two-year
6:40
pilot period only
6:43
to demonstrate some success. Wow. Things
6:46
have changed quickly. Things have changed pretty quickly. And
6:48
they haven't just changed in our field. There have
6:50
actually been a lot of other groups
6:52
that have suddenly got interest in this too for
6:56
the same reason that basically the
6:58
public and almost every
7:00
scientist or technically-minded person is now
7:02
thinking about ways they could possibly
7:04
apply AI to that work. The
7:07
tools are just so much more powerful
7:10
that the computational technology
7:12
is powerful. And I
7:15
think we've all seen how transformative it can
7:17
be in some examples
7:19
like large language models. So
7:21
what were some of the limits or I guess I should
7:23
say are some of the limits of
7:26
FORTRAN based modeling and
7:28
how are you able to overcome them or
7:30
might be able in the future to overcome
7:32
them using machine learning? Yeah, well, many of
7:34
those limits don't actually have to do with
7:36
FORTRAN, but they have to do with the
7:39
algorithms that we use to, for
7:41
instance, represent the flow of air
7:44
and the atmosphere, the wind basically.
7:47
They basically have to do fundamentally with
7:49
how you numerically solve
7:52
equations on grids. And
7:54
without describing any details, basically they require
7:57
a limit on how fast a model
7:59
like that can be stepped forward and
8:03
what size stops it can be stopped.
8:05
And for a climate model, that means
8:07
that climate models written and phrased
8:09
in the traditional way require time
8:12
steps of a few minutes. With
8:15
machine learning, we are able to
8:17
actually sidestep that limitation because our
8:19
algorithms are not formulated in the
8:22
same way. And so
8:24
we can take much longer steps of many hours.
8:26
And that by itself allows the
8:29
machine learning to accelerate the model
8:31
by a factor of 20 to
8:33
50. Wow, okay. Even
8:36
with no improvement in computation. Right. Furthermore,
8:39
machine learning runs very efficiently
8:41
on GPUs. And
8:43
so we're able to much more efficiently use
8:45
the computational resources that are
8:48
available. Right. And so
8:50
that actually adds to
8:52
the efficiency yet further. I
8:54
don't know if this is a good question
8:56
to ask per se. So tell me if
8:59
I'm off base here. But can you compare
9:01
the size of a climate model to something
9:03
like a large language model just to kind
9:05
of give the offer me but
9:08
also to give the audience a sense of, you
9:10
know, in my head, describing
9:13
the wind in an algorithm seems much
9:15
more complicated than describing a sentence of
9:17
text. But I really have no idea. Yeah.
9:21
And of course, what we're not we're
9:23
not comparing describing the globe
9:25
to describing a single word or single
9:27
sentence, we're sort of thinking about it,
9:30
describing all the sentences that you might
9:32
make. And so it's
9:34
more like, you know, the
9:37
encyclopedia Britannica versus representing the
9:39
Earth on a global grid. And of course, an
9:42
issue is that it depends on how detailed
9:44
the grid is. If you
9:46
represent the earth on a hundred
9:48
kilometer grid, you might be talking
9:50
about hundreds of thousands
9:52
of different grid points and parameters that
9:54
you're trying to advance in a model
9:56
at each time step. If you're
9:59
talking about a kilometer scale model
10:01
of the kind that we're actually
10:03
hoping to train our machines learning
10:05
on, you're talking about
10:08
hundreds of millions to a
10:10
billion different points. But
10:12
that's still small compared to the number
10:15
of parameters in a
10:17
very large language model where there might be
10:19
a trillion parameters in where you're definitely training
10:21
on trillions of tokens of
10:24
data. And so when we're talking about
10:26
a climate model, how does
10:28
it work kind of for the
10:30
layperson to understand, is it a
10:32
single model of the Earth? Are
10:34
there different models for different geographical
10:36
regions? Is it something totally different?
10:39
How do these models represent what
10:42
someone like me might think of right, like
10:44
the planet and like climate change over time?
10:48
So we modularize the model
10:50
into different components representing parts of
10:52
the climate system. And in fact,
10:55
part of the climate system we're
10:57
working to improve by machine learning
10:59
in our project is the atmosphere.
11:02
So we represent the entire three
11:04
dimensional structure of the global atmosphere
11:07
by breaking it into about a
11:09
hundred layers and representing
11:13
each of those layers on a horizontal grid.
11:16
There are also somewhat
11:18
similarly constructed models of the
11:20
ocean for representing ocean currents,
11:22
which are much
11:24
slower but are complicated too and
11:27
have a lot of detail. And
11:30
those models interact with each other
11:32
through what's called a coupler, which
11:34
allows information to be exchanged between
11:36
these models. Similarly, there are
11:38
models of
11:41
ice, of sea ice, for instance, of
11:44
chemical processes, such as
11:48
create the little particles
11:50
that actually cause
11:52
clouds to nucleate when
11:54
water vapor condenses And
11:57
a whole variety of other processes. So.
12:00
Climate model consists of all of these.
12:02
Are. Different components interacting through
12:05
on. A couple are some
12:07
time. It's kind of the it's if
12:09
you want the conductor as the orchestra,
12:11
freight rates and then what are you
12:13
able to see from the model? What
12:15
kind of information is it for testing?
12:17
Is it current conditions? Is it what?
12:19
What kind of information are you getting
12:21
back from the law? So to think
12:23
about it's a traditional climate model and
12:25
actually it but were trying to achieve
12:28
and machine learning as being. An
12:31
Earth System simulator? So it
12:33
is. I'm it's trying to
12:35
predict climate. and climate is
12:37
by definition. Ah, you know
12:39
how. How to the
12:41
long term conditions look like? How
12:43
are. Things. Like out. Extremes.
12:46
Of precipitation Changing that it's
12:48
statistics of climates is a
12:50
way it's getting at that
12:52
is it's actually simulating. The
12:54
system is basically making artificial
12:56
weather and the ocean equivalents
12:58
in a changing environments. Ah
13:00
that allows the different model components to talk
13:03
to each other, the atmosphere to talk to,
13:05
the ocean to talk to us, the biology
13:07
and all the others, The chemistry, insulin, Or
13:10
but it's a simulator And so as
13:12
a result, when we do machine learning,
13:15
we actually retain that start of trying
13:17
to make a simulator which is just
13:19
predicting a mirror on artificial or pretend.
13:22
Likely our evolution of how atmosphere month
13:24
and yes might go and I'm I'm
13:27
a little hesitant or perhaps ah circles
13:29
the asked by out what are the
13:31
models saying what is what is the
13:33
future look like for for all of
13:35
us on earth right? Well I a
13:38
I conventional climate models have told us
13:40
everything we need to know s to
13:42
know that's ah you know the earth
13:44
is getting warmer. That has many consequences
13:47
which a very robust warms nora land
13:49
at Worms especially in the polar regions
13:51
basically in order to. Turn that around.
13:53
Regained has to get to a carbon
13:55
neutral power system and economy as fast
13:57
as we town. The effects of emitting.
14:00
The dioxide a cumulative so we know that
14:02
and we have all the information and which
14:04
we need to act and so you might
14:06
ask. Okay, well, why do we need climate
14:09
models? It's there Any better than that. And
14:11
the reason is that you know as you
14:13
think about society and how we generate. Our.
14:16
Power. ah it's kind of my cab.
14:18
It's a huge battleship. It has many
14:20
components and still it's also like a
14:22
balance it without one captain. As as
14:24
hundred sit ups and semester the ships
14:27
they have a lot of other priorities
14:29
undermined other than climate change. no matter
14:31
how hard they are how much they're
14:33
aware that that's in this year and
14:36
so so we are going to be
14:38
dealing with a warming climate and all
14:40
the yes extreme heat and more extreme
14:42
precipitation saw on that is going to
14:44
creates we are going. To have to
14:46
plan thought were going to have to adapt
14:49
to that as well as try to mitigate
14:51
it by switching away from. I.
14:53
Greenhouse dependency on and so
14:55
it's not as adaptation that
14:57
we made the able to
14:59
inform better with better climate
15:01
models because for instance if
15:03
you. Say.
15:05
Are are a homeowner in the
15:08
suburbs of Allies eyes enough to
15:10
an area which was recently as
15:12
affected by atmospheric river that was
15:15
quite destructiveness. Ah you are wondering?
15:17
Okay well in my water said
15:19
are we likely to see in.
15:22
Rainfall. Events that are twice destroying thirty
15:24
years from now. And if so, what do I
15:26
have to do for a new How do I?
15:28
dogs? Are. How to we as
15:31
a society build our infrastructure or
15:33
keep people from loving in certain
15:35
places or plant different crops in?
15:37
All Those things are very local
15:39
decisions and so for many of
15:42
those decisions we need very local
15:44
information. Because climate change might
15:46
be, you know it might affect the
15:48
windward slopes of Hawaii. Different. An ugly
15:50
words. First phone since his house I'm
15:53
so that's where I. That's why we
15:55
really can help to do better to
15:57
is localized modeling. One of the things
15:59
that. The top of the less
16:01
that machine learning eyes able to help
16:03
us? Yeah, so of that's that's certainly
16:05
part of our vision. So the idea
16:07
really is is two fold with what
16:10
we can do with machine learning. The
16:12
first part of it has to do
16:14
and localized modeling but rather and directly
16:16
And that is that. Climate models
16:18
are typically run on a grid
16:20
as I'm sixty two one hundred
16:23
kilometers or south in. The reason
16:25
to that is that. You.
16:27
Have to be able to compute for
16:29
hundreds of years. So you need to
16:32
be able to simulate the earth for
16:34
hundreds or thousands of years with a
16:36
model. You can computationally afford to do
16:39
that. On the other hand, we can
16:41
simulate the earth much better with a
16:43
model with a much finer grads say,
16:45
two or three kilometers sept two miles
16:48
or something like that's that kind of
16:50
his neighbourhood scale. practically. nice. And so
16:52
it represents all the details of the
16:55
earth's surface. and ah, and the atmosphere
16:57
and individual. Storm Systems and South's and
16:59
sometimes it's a fundamentally better model. So
17:01
if we can train or machine learning
17:03
it on models like S which we
17:05
can't afford to run for hundreds of
17:07
years. But. The machine learning
17:09
version or emulate or of the model
17:12
ten run for hundreds of years. Many
17:14
times we can get the information with
17:16
a lover's information that have that find
17:19
good model to help a and or
17:21
productions such one of the goals. But
17:23
the other thing is that nevertheless when
17:26
we train the machine learning model we
17:28
don't train it on a very fine
17:30
bread. Me trying to have actually typically
17:33
under fifty or hundred kilometer grads but
17:35
then another machine learning tool called Jenner
17:37
Dovey I'm particular. something called the season
17:40
but lengths can be used to take
17:42
the information we have i done hundred
17:44
condor scale and downscale at noon two
17:46
or three kilometer model to scale with
17:48
the help of this find good model
17:51
the we used for training and and
17:53
so these two parts kind of sit
17:55
together lights hand in glove it's existence
17:57
of these time for models allows
18:00
us to develop ML
18:02
to emulate them and then also allows
18:04
us to downscale that information back to
18:06
that very fine scale which people wanted
18:08
to help people in figuring out
18:11
what to do with that bridge in the neighborhood that
18:13
they're afraid is going to wash out. Right.
18:17
You took me back to my own neighborhood. They're doing
18:19
work on the bridge right now, in fact. Yeah, right.
18:22
So, I have two questions. One is
18:24
sort of about your work
18:26
specifically, but before that, since you
18:28
mentioned talking about, we're talking about
18:30
localized modeling and things that people
18:32
and your neighborhoods and local governments
18:34
might be able to do with
18:36
some of the forecasting information, what
18:39
if anything, can individuals, can
18:42
small communities be doing now
18:44
to prepare for what climate
18:47
change has in store for us in
18:49
the near and kind of medium future? Right. Well,
18:52
I mean, I think the first thing to realize is every
18:55
culture, every region, everywhere is
18:57
going to have to change
19:00
the way that it lives in
19:02
many ways. We have to move to EVs,
19:04
but we have to generate concrete. We
19:07
have to make concrete differently. We have
19:09
to farm in ways that release less
19:11
nitrogen oxides to the atmosphere. It's another
19:13
greenhouse gas. We have to release less
19:15
methane from fracking. And so, this is
19:18
a problem where we have a death
19:20
by a thousand cuts. And so, don't
19:23
pretend that someone else's technology is going
19:25
to solve the problem for you. But
19:28
this is a problem where we
19:30
have to consistently support
19:32
policies that take
19:34
a holistic view of the problem.
19:37
Carbon taxes, policies that
19:39
basically say, you know, we will
19:41
emit 50% less CO2 by 2035 than
19:45
today. If we don't do that,
19:48
we're not going to get to the end game. And
19:50
so, this is an urgent problem. What we
19:52
do now matters a huge amount for the
19:54
future. So that's one thing for people to
19:56
realize. And we know everything about that right
19:59
now. And I can't
20:01
stress more how important it is to
20:04
work on the prevention, because
20:06
that is in the end the way we'll get to
20:09
the end of this problem. But then
20:11
thinking more about, okay, and what will
20:14
more AI do for people's view of
20:16
this problem? As I said, we
20:18
are going to have to adapt. And
20:20
the role of AI here is to
20:24
basically make the amount of computational
20:27
power and the technological prowess you
20:29
need to have in order to
20:31
get a climate prediction, climate
20:34
change prediction for your backyard, it's
20:36
gonna bring it down to the
20:39
scale where individual communities potentially can,
20:42
I won't say they can run their own model, but
20:44
they can find someone who can do it for a
20:46
reasonable price using cloud
20:49
resources, using things that
20:51
are within their capability to go
20:53
after. And so it will
20:55
be the case that no one will be able
20:57
to afford or want to make
20:59
planning decisions without accounting for
21:01
climate information. And
21:04
right now there's a lot of friction in that process,
21:07
because there's a long way from the
21:09
huge climate modeling centers somewhere else to
21:11
the decision you have to make next
21:13
week, but that doesn't have to be
21:15
the case. And really, ML
21:17
is a way to get there. And
21:20
a feature of ML that's particularly important
21:22
there is that I
21:24
think we'll be able to combine these
21:26
kinds of models that will tell you about
21:29
climate with language models so that you can
21:31
even query them with the questions you
21:33
wanna ask and the way you want to ask them
21:35
and they'll be able to give you sensible answers. That's
21:38
really the sort of biggest vision here. Yeah, and
21:40
so kind of along those lines, I was gonna
21:42
ask what's next
21:44
for the work that you're specifically doing,
21:46
or maybe kind of fold it into
21:49
that, are there problems
21:51
right now from kind of a
21:53
science and technology standpoint that
21:55
you're working on that,
21:58
not that you're waiting for... for someone else's advancement to
22:01
help you get over the hump. But
22:03
are there kind of hurdles right now that you're
22:05
trying to get past that, you know, can kind
22:07
of pave the way for the next part of
22:10
the vision? Right,
22:13
so there are sort of two different kinds of
22:15
hurdles. One of them is I mentioned that the
22:17
earth system is complex and the way that we
22:20
model it is modularized. And
22:23
in machine learning, we
22:26
find that you can get a much
22:28
better answer if you can put
22:30
your entire model under the machine
22:32
learning hood and sort of optimize
22:34
everything at once. Sure, right. But
22:37
when you have a very modular starting
22:39
model and some of the components are
22:41
not available in some kind of machine
22:44
learning form, that's just to be
22:46
difficult. So one thing
22:48
we're trying to do is work with our
22:50
other colleagues
22:53
doing machine learning, say of
22:55
the ocean, to
22:57
try and at least build emulators of
22:59
the ocean, couple them to
23:01
emulators of the atmosphere, try to have
23:04
them emulate sort of coupled atmosphere ocean
23:06
models, because that's the minimal configuration that
23:08
I think we need in order to
23:12
talk about 21st century
23:14
climate change. So this
23:16
coupling of the atmosphere and the ocean is one
23:18
big challenge that we're working on right now with
23:21
colleagues like Laura Zana, who's
23:24
talking later in this GTC. And
23:28
forgive my ignorance in asking this, but
23:30
is that a software problem? Is it
23:32
a sort of infrastructure and hardware problem?
23:34
Is it a, you know,
23:37
humans kind of getting together and figuring
23:39
out how to best work together problem?
23:41
It's a, you know, it's a people
23:43
power problem in that machine learning software
23:45
doesn't design itself from scratch. And
23:49
an emulator of either the atmosphere
23:51
or the ocean is
23:54
a piece of software which starts
23:56
with something you might be
23:58
able to take from another field. like
24:00
a vision model or a video
24:02
model or something like that. But
24:07
it has to be adapted to your particular
24:09
situation and that takes experts and it takes
24:12
a lot of decisions and it
24:14
actually takes a lot of
24:16
playing around with the model. And so
24:18
it's really people power. So in this
24:20
particular case, we have an
24:22
atmospheric emulator from AI2 that
24:25
I think is an appropriate atmospheric component
24:27
for such a model. But
24:29
the oceanic equivalent doesn't really exist yet.
24:32
An ocean that you can play for
24:34
a long time in this emulator mode
24:36
and have it stay stable and
24:39
accurate and represent meaningfully
24:41
a climate. And so
24:43
that's a real, it's
24:46
a development challenge to sort of a
24:48
mixture of science and technology. It's not
24:51
really being limited by the hardware that
24:53
we have right now. But
24:55
once we start running this model, we'll
24:57
get limited by the hardware because we'll
24:59
be wanting to run the emulator for
25:01
hundreds and thousands of years. And
25:04
I should mention that actually our
25:06
climate model is based on the
25:08
SFNO architecture of
25:11
a product called ForecastNet, which
25:13
an open source piece of
25:15
software that was developed in
25:17
video. And SFNO and developed
25:20
by Boris Bonov here has
25:22
been totally instrumental in making
25:24
our project successful. But
25:27
another thing that we're also trying
25:29
to start collaborating with NVIDIA on
25:31
is this downscaling issue. How do
25:33
you take information from an
25:35
emulator that might work natively at 125
25:37
kilometer scale and
25:41
really reduce it accurately to a
25:44
local scale using machine
25:46
learning based statistical techniques that are
25:48
very efficient and that we can
25:51
sort of apply on
25:53
as little as a single GPU. And
25:55
so that's also a challenge that
25:57
we're currently working on. How do
25:59
you... measure the accuracy
26:02
is the right word, the kind of the
26:04
performance of an emulator
26:06
that's doing something like climate modeling
26:09
out into the future? Yeah, that's
26:11
a great question. So when people
26:13
have developed weather emulators, which have
26:15
actually weather forecast emulators have been developed,
26:18
and at this point, they're actually more
26:20
skillful than our most conventional weather forecast
26:22
model in the world after
26:24
10 days, which is remarkable since
26:26
these emulators have only been developed
26:28
for the last few years. The
26:30
emulators are better at that. They
26:34
make better forecasts than the original models. The way
26:37
that they do it is they're trained on the
26:40
last 50 years of historical
26:42
data blended together using these
26:44
physically based forecast models. So
26:47
they still are actually still relying on the
26:49
physically based model for
26:52
their initial guess. But from then on, they
26:54
can make forecasts that are
26:56
more skillful simply by
26:58
training on observed data. Now
27:01
the problem is if you take these same
27:03
models and you try to run them out
27:05
for longer, a lot of them drift away
27:07
or they go unstable. They're not actually suitable
27:09
for climate directly. But nevertheless,
27:12
so one obvious way to train a climate
27:14
model would be try and train it the
27:16
same way. But the only problem is the
27:19
future is not the present, is not
27:21
the past. And one
27:24
of the adages of machine learning is
27:26
you should never expect your model to
27:28
generalize too much out of sample. So
27:30
we don't do that. And instead, our
27:32
approach to emulation is we
27:35
are just going to try and emulate physically
27:37
based climate models. So right
27:39
now, the physically based climate model is one that
27:42
a major climate modeling center could
27:44
affordably run for themselves. So we're
27:46
doing nothing that is of interest
27:48
to them, except that we can
27:51
do it many, many times faster. But
27:53
in future, our goal then, as I
27:55
mentioned before, is to emulate
27:57
these three kilometer global models, which are
28:00
are inherently more accurate and believable, not
28:02
just in the present climate, but also
28:04
in future climates. And
28:07
at that point, we're doing something which you can't do
28:10
right now. And
28:12
so that's sort of the
28:15
aim. But nevertheless, it
28:18
is important to recognize we're just trying to
28:20
emulate another model. So we inherit all
28:23
of the pros and cons of
28:25
the other model. You're still extrapolating,
28:27
but now you're taking advantage of
28:29
the physically-based model to
28:31
use the physical reasoning
28:34
and physical laws that we can build into
28:36
that. And we're using
28:38
that then to train
28:40
our machine learning in multiple climates
28:42
so that it can also do
28:44
that same job of being able
28:47
to predict the cross-changing climates. Right.
28:49
Are there certain regions where it's harder
28:51
or easier to forecast
28:53
the climate and climate change than
28:55
others? Yeah, actually, regions
28:58
like California are quite difficult.
29:00
I think you could ask
29:02
hard versus, well,
29:05
what's the problem? So in
29:07
some sense, the problem
29:10
in places like the
29:12
Southwest US is there's
29:14
a lot of variability, but it seems like
29:16
the climate is getting drier. And is it
29:18
getting drier? And it turns out if you
29:21
query something like 20 or 30 of
29:23
the world's leading climate models
29:26
today about that, they don't all
29:28
agree. And so some
29:30
of them say, okay, the Southwest is drying
29:32
out. Others of them say, well, it's not
29:34
really drying out. It's not getting any wetter,
29:36
but it's not getting any wetter. And so
29:38
there are regions of the world, and especially
29:40
these are kind of in the edge of
29:42
the tropics in regions sort of
29:45
at the latitude of California, where it's actually hard
29:47
to predict what some of the trends are going
29:50
to be like. And
29:52
which are also subject to a lot of extreme
29:55
events. And so those
29:57
actually are complicated regions. issues,
30:00
though, are things like there's
30:02
some areas which are very
30:04
susceptible to, say, tropical cyclones,
30:06
hurricanes and typhoons, where
30:09
one extreme event can wipe the place
30:11
out. And so they care a
30:13
lot about something that's very rare and not
30:16
so easy to simulate. And so in some sense,
30:18
those are also hard areas to
30:20
forecast. Right. That makes
30:22
sense. Chris, for listeners who would like to
30:25
find out more about climate modeling, about the
30:27
work you're doing, about the work AI2 is
30:29
doing more broadly, where
30:31
would you direct them to look online? Well,
30:33
we have a website. So
30:36
AI2 has a website,
30:38
lnai.org is our name.
30:42
On there, there is a link
30:44
to climate modeling that talks about
30:46
our group's activities. I should
30:48
emphasize, though, that again, there are many groups
30:50
other than our own that are now working
30:52
in this area. And so hopefully
30:54
you'll read about this in places like The
30:57
New York Times, The Economist, and Washington Post,
30:59
which already actually have been doing a very
31:02
good job of covering
31:05
the explosion of weather forecast emulators and how
31:07
they're transforming the field. So I'm thinking you'll
31:09
read about this area of research in the
31:12
news in the not too different future as
31:14
well. I hope so. And I hope it
31:16
moves us to act. Chris
31:19
Brutherson, thank you so much for taking the
31:21
time out of GTC to join the podcast.
31:23
Best of luck with everything you're doing and
31:25
your colleagues. We're all, I was
31:27
gonna say we're all counting on you, but that's a lot of pressure
31:29
for a podcast. But we appreciate the
31:31
work you're doing is an understatement. It's vital. Okay,
31:34
well, thanks very much, Noah. It's a pleasure to be here. www.mooji.org
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More