How AI helps predict extreme weather | BBC News

00:00

we are back with our weekly segment AI

00:04

[Music]

00:06

decoded welcome to the program we have

00:09

had a summer break from AI decoded but

00:11

if like me you were on the British

00:12

beaches Sheltering from the rain then

00:15

maybe you were scaning your mobile

00:16

weather app to see if the sun might ever

00:18

reappear which got us thinking what

00:21

about Ai and the weather how do you

00:23

predict climate when it is changing so

00:26

fast how do you process that incredible

00:28

amount of computerized data that is now

00:31

being generated well you model it and

00:34

that is where AI is making huge advances

00:37

there is a forecasting Revolution

00:39

underway so accurate says the guardian

00:42

and now in much more accessible format

00:45

that very soon governments around the

00:46

world will be able to save lives and

00:49

protect livelihoods before extreme

00:51

events even occur we'll hear from the

00:53

team at Oxford University who are

00:55

filling in the gaps with AI and making

00:57

it more readily available through Cloud

00:59

computer shoting or how about this from

01:01

the EU destination Earth a digital copy

01:05

of our planet on which scientists are

01:07

running complex simulations to predict

01:09

natural phenomena AI combined with

01:12

climate science powered by

01:15

supercomputers a digital twin if you

01:17

will that will help scientists predict

01:20

the evolution of climate change with me

01:23

has ever our regular commentator and

01:26

colleague uh Stephanie har is here also

01:29

in the studio the very well-known

01:30

meteorologist Florence rabier Dr rabier

01:34

director general of the European Center

01:36

for medium-range weather forecast and

01:38

joining us also on Zoom Professor Steven

01:40

Belcher who is chief of Science and

01:42

Technology at the UK Met Office welcome

01:45

to you all um fla uh we're going to

01:49

start with you um and the Earth's

01:52

digital twin that you and your team have

01:54

built in collaboration with the the AI

01:57

industry so let's get a view for the

02:00

viewers let's let's just show the

02:01

viewers what it entails and we'll talk

02:03

off the

02:06

back to create a better future we must

02:10

push the boundaries of

02:13

today simulations of our Earth system

02:16

known as digital twins will help us

02:19

understand predict and plan for a

02:21

rapidly changing

02:25

world this twin will offer highly

02:28

detailed Interactive

02:31

data that can support decision making

02:34

around extreme weather

02:38

[Music]

02:48

events this twin will show us possible

02:52

Futures it simulates different climate

02:55

change scenarios over many decades

03:01

helping us to be ready for whatever

03:07

happens wow so in simple terms you are

03:13

stimulating the natural phenomena and

03:16

the human activities on Earth putting it

03:18

all together through this supercomputer

03:20

and what putting it

03:22

onto a digital twin of the earth yes so

03:26

digital twin of the Earth because it's

03:28

this super model of everything that the

03:31

Earth is doing that we can predict

03:33

through Computing equations so it's a

03:36

model where all our knowledge of the

03:38

physics of the atmosphere and the Earth

03:41

system is encompassed in that model so

03:44

it's a computer program that we put on a

03:46

supercomputer and we run it but all our

03:49

knowledge of the physics accumulated

03:50

since Newton and all is there about

03:54

gravity condensation you know storms Etc

03:58

and it's a digital twin of the Earth

03:59

because it's very accurate and it has a

04:02

very high resolution and also it's

04:04

interactive you can play a bit with it

04:07

and simulate what if scenarios so I

04:11

imagine that in Times Gone by you would

04:13

do that at a very local level but of

04:16

course we're all interconnected the

04:18

world is is is a global environment our

04:22

our weather systems uh and activities

04:24

are all connected so so how does this

04:27

enable you to improve the forecasting

04:29

that you do yeah so at ecmwf what we do

04:32

we run these Global model so it's really

04:35

across the whole world because if you

04:36

want to know the weather in Europe now

04:39

you have to know what happened in the US

04:40

a few days ago and in the Atlantic and

04:42

even in the Pacific if you want to

04:44

predict the weather 7 Days beforehand so

04:47

it's all interconnected you're right you

04:48

have to start with a global scale and

04:52

then you can refine at the local scale

04:54

as well but you really have to know what

04:56

whatever happens on the world at any

04:59

point in in time in order to go further

05:01

in time in your predictions and that's

05:03

what we've been doing for about 50 years

05:06

in collaborations with our member states

05:09

35 countries in Europe supporting this

05:11

work so Florence is this a uniquely

05:13

European initiative or do you work with

05:15

other partners around the world so there

05:17

are several Global models in the world

05:19

and our because we are European we are

05:22

working in collaboration with 35

05:25

meteorological services in broadly

05:27

Europe broad Europe but Amer model Etc

05:31

ah okay but are you using historical

05:34

data or like live data or both probably

05:37

both I mean for the weather forecast

05:39

usually we use the current data so the

05:42

data we've seen in the last 12 hours but

05:45

the model had seen the data beforehand

05:47

and it's a sort of continuous process we

05:49

combine physics and data we go forward

05:51

combine again so the latest forecast

05:54

uses the latest data but historically

05:57

we've used the data from the last

06:00

decades and rolling like this I tell you

06:02

what we we've got a a real life example

06:05

uh that mapped the recent typhoon

06:07

typhoon G I think it was called uh in

06:10

July this year so so you see all the

06:12

lines around the two main lines so

06:15

there's the red line and the black line

06:17

which we'll talk about in a second what

06:18

are the other lines that we're looking

06:19

at so that's typically what we do as a

06:22

forecast so we we predict the weather

06:24

but in particular we concentrate on

06:26

severe events like that because this

06:28

typhoon G me was really uh had dire

06:31

circum you know consequences with 100

06:33

people dead and millions affected so

06:37

what we produce every day we produce not

06:39

just one forecast but we produce

06:41

actually several forecast together to to

06:44

depict the whole probabilities of what

06:46

the weather will do this way we don't

06:48

just simulate the track of the typhoon

06:52

but all the possible tracks that we

06:54

think the typhoon will take in the next

06:55

few days these are the gray lines so

06:58

which is the AI model

07:00

so the gray Line's our physics based

07:02

model and the black line is the real

07:05

observed track of typhoon and in the

07:08

blue what you have is our best estimate

07:12

that we had before AI of where the

07:14

typhoon would go and would hit China

07:16

he's telling me that red one is the AI

07:19

that's almost tracking exactly what

07:20

happened in real time in that case it is

07:23

and that's incredible it is incredible

07:25

but you can't judge everything on one

07:27

case of course we accumulate all the

07:29

these cases and we do statistics but it

07:31

is true that the AI models are in

07:34

general about 25% more accurate in

07:38

predicting the track of tropical

07:40

Cyclones typhoon and hurricanes which is

07:43

huge but they are not doing everything

07:46

right either so in particular in terms

07:48

of the intensity of the typhoon they are

07:51

actually about 20% worse so it's not all

07:54

perfect you've got a real interest in

07:55

this because I know that you grew up in

07:57

tornado Valley in America right so

08:00

Tornado Alley is the midwestern part of

08:02

the United States that runs from North

08:03

Dakota all the way down to Texas and

08:06

then probably for 500 to a, miles on

08:08

either side so I grew up just outside

08:10

Chicago and routinely we practice these

08:13

drills as children you get a little bit

08:16

of warning we're talking seconds and you

08:18

have to find the nearest basement and

08:19

get underground because a storm will

08:21

come through that can destroy an entire

08:23

town in seconds presum this could tell

08:25

you which street to go to well we're not

08:28

at that sort of scale especially as we

08:30

work I have big

08:32

expectations but but it's get but it's

08:34

getting better and better all the time

08:36

it is getting better all let's bring in

08:38

Professor Steven Belcher who is the

08:39

chief of Science and Technology at the

08:41

UK Met Office um welcome to you um so

08:45

far we've talked about global weather

08:47

patterns uh Stephen climate Trends

08:49

mapping evolution of weather patterns

08:51

but how much more precise is weather

08:54

prediction getting dayto day because of

08:56

this AI

08:57

technology well it's worth remember

08:59

remembering that there's always pressure

09:01

to increase the accuracy and the utility

09:03

of weather forecasts uh today's great

09:06

example we've had certainly lots of rain

09:08

here down in the southwest of England so

09:10

with climate change making extreme

09:12

events even more extreme we're demanding

09:15

that our forecasts get better to uh help

09:18

us understand what the impacts of those

09:20

might be also we've got new applications

09:23

of weather forecasts and just think

09:25

about the roll out Renewables this is

09:28

meaning that weather is now the fuel of

09:30

the future so understanding that fuel is

09:33

another application of of our weather

09:36

forecasts and so to make them more

09:38

accurate we need increased lead time so

09:40

we need warnings further ahead of when

09:42

we're getting these extreme events but

09:44

we'd also like finer detail just as

09:47

Florance was describing earlier I was

09:50

making quite big demands of Florance but

09:52

I'm going to put up an image here about

09:54

just how accurate this can get so here's

09:56

a here's an image of London that we'll

09:58

all recognize um uh you can see the

10:00

Millennium D there the bend in the River

10:03

East London tell us what we're looking

10:05

at here on the left and on the

10:08

right right so at the Met Office we

10:11

complement What's Done in Florence's

10:13

organization the European Center by

10:16

producing highresolution physic Bas

10:19

modeling of the weather over the UK so

10:22

the leftand side is showing you the grid

10:25

that we divide London into in order to

10:28

provide that weather so what what we're

10:31

seeing now is the weather on at that

10:33

kind of resolution about 1 and 1 half

10:35

kilom we we kind of increment the the

10:39

differences in the rainfall and the

10:41

temperatures what we've been able to do

10:43

in fact one of our Rising Stars Lewis

10:45

blun here at the Met Office working with

10:47

students at the University of reading

10:49

and also at the Bureau of meteorology in

10:52

Australia um has devised a machine

10:54

learning techniques to add fine detail

10:57

onto those routine for CS that gives us

11:01

detail resolutions of hundreds of meters

11:04

in the temperature so you can tell the

11:07

temperature and the Heat and the

11:08

rainfall literally over the Millennium

11:10

Dome uh the temperature at this stage

11:13

and maybe other variables in the future

11:16

but but the the temperature is what

11:17

Lewis and his colleagues uh worked on

11:21

and the reason this is important is that

11:22

we we've known for some time that uh

11:25

when we have heat waves um those heat

11:27

waves are more extreme in urban areas so

11:29

in particular those who live in cities

11:31

will have noticed that the temperatures

11:33

don't cool down so much at night and we

11:36

call that the urban heat island so what

11:39

Lewis and his team did was to take data

11:42

actually from

11:43

crowdsourced data in back Gardens and uh

11:48

citizens in London their data so of

11:52

variable quality frankly plus five

11:55

professional weather stations they Mash

11:58

that together together with machine

12:00

learning tools and augment the regular

12:02

forecast that we produce here at the Met

12:05

Office and can then produce these

12:06

temperature forecasts at these very fine

12:08

levels so it's another example of how AI

12:12

can really change what we're doing in

12:14

the in the weather World Professor beler

12:17

I've got a question for you exactly

12:19

about this um taking data from your back

12:21

Garden sometimes when I'm standing in

12:23

London I will consult the Met Office app

12:26

regularly in fact and it will say that

12:28

it's sunny and I'm being rained on why

12:31

is that happening and second how or when

12:34

will I be able to send data to you

12:36

saying no no here in hatney it's

12:40

raining so you can you can send data to

12:42

us right now so it's called weather on

12:44

the web the wow site so please do that

12:48

you can Lodge your longitude and

12:50

latitude and send in your data to us and

12:52

as I say that's the data that Lewis and

12:55

his colleagues used to produce this um I

12:58

think

12:59

in terms of weather forecast let's not

13:01

forget that over the last 50 years

13:03

through the Advent of satellite

13:05

observations and other observations

13:07

right around the world the Improvement

13:09

of those global models that Florence was

13:11

describing earlier and and the increased

13:14

scale of supercomputers that we've got

13:16

these physics-based models we've got the

13:18

weather forecast has increas improved

13:21

tremendously well and and one of the

13:23

statistics we describe is that the

13:25

weather forecast incre improved by one

13:28

day per decade so the 4-day forecast now

13:33

is as good as the three-day forecast was

13:36

10 years ago so this is often referred

13:38

to as the quiet revolution in weather

13:40

forecasting what AI is doing is really

13:43

accelerating that Revolution so it's a

13:45

loud Revolution we're going to continue

13:47

the conversation we have to get a short

13:49

break but we're going to see how this

13:50

can be applied around the world um

13:53

coming up after the break we'll bring in

13:54

the climate physicist Dr shuy Nath she

13:57

is part of the physics team at Oxford

13:59

where they just pioneered a new approach

14:01

to predicting extreme wither stay with

14:03

us welcome back we are warned repeatedly

14:06

that climate change will affect millions

14:08

of people worldwide in fact is already

14:10

affecting lives and livelihoods and

14:11

particularly so in some of the poorest

14:13

regions of the world where they don't

14:15

have access to this realtime forecasting

14:17

or the vast computer power needed to

14:19

produce it well Dr shuy na is a climate

14:23

scientist at Oxford University she's

14:25

been working with the UN world food

14:26

program to develop an AI system that is

14:29

pulling together all this data current

14:31

and historic and applying that to

14:34

localized area that information can now

14:36

be condensed and shared through cloud

14:38

computing to help the governments and

14:41

Aid agencies better prepare for climate

14:44

disasters let's bring in our guest then

14:46

Dr shudy D It's good to talk to you

14:48

thank you very much for coming on the

14:50

program

14:52

um there were some brilliant physicists

14:54

like you in the in the Oxford University

14:56

Department I I want to better understand

14:58

that what the AI is doing to speed up

15:01

the process and fill in the

15:04

gaps yeah so thank you for having me at

15:07

oxid physics we're exploring hybrid

15:10

modeling approaches so we're looking at

15:12

how AI can best complement our existing

15:15

physical weather models so as Florence

15:18

says these models have all the physical

15:20

knowledge that we've accumulated since

15:23

Newton and we're complimenting it with

15:26

AI particularly at Oxford we're looking

15:30

at rainfall since this is a high impact

15:33

very localized feature of the weather

15:36

and what we're seeing is that when we

15:39

take the best of the physical weather

15:42

forecasts um we can really use AI as a

15:46

data driven technique to correct the

15:48

structural errors that exist in these

15:50

physical forecasts um that could arise

15:53

from incomplete representation of

15:55

atmospheric processes to better deliver

15:58

actual accurate rainfall forecasts um

16:01

within the region that we work in so

16:03

that's the greater Horn of

16:04

Africa and um Dr n how do you see

16:09

Ordinary People in the regions where

16:10

you're working being able to access this

16:12

very sophisticated and high powerered

16:14

technology that you're working

16:16

with so that's a very good question and

16:19

that's actually in my opinion one of the

16:22

real strengths of AI it's a very lowcost

16:25

lightweight model of being able to

16:27

represent very complex phenomena so we

16:30

work closely with all the local

16:32

mological bodies and we work with

16:35

developing the model with them so they

16:37

actually run the AI models um inhouse

16:42

and that means that they can actually

16:44

generate weather forecasts on a laptop

16:47

mind you that's a laptop as compared to

16:50

a supercomputer which is what typically

16:52

is used to generate weather forecast can

16:54

you give us an example of where you've

16:55

used that a real a real time example of

16:59

course yeah so we use it in Kenya so

17:01

they update the forecasts every day um

17:04

in Kenya and Ethiopia and the forecasts

17:07

are also available on a website so um

17:10

the website name is sean. pac. net and

17:14

they're updated every day um from the

17:18

in-house forecast generated on you know

17:21

their uh equipment so it really is a way

17:24

of giving these people a bit more access

17:28

ible weather forecasting but presumably

17:31

that I mean the Breakthrough of all this

17:34

is that you can if you know what's

17:36

coming and and the long range

17:37

forecasting improves you can the aid

17:40

agencies can can put can sort of store

17:44

forward the aid that they're going to

17:46

need for what's coming at them so you

17:49

know so often on our programs we're

17:51

we're sort of saying well we can't get

17:53

to these inaccessible areas but now the

17:55

stuff will already be there because

17:56

we've already forecasted what's coming

17:59

exactly so we particularly actually

18:03

focus on linking research to actions so

18:06

we work with linking these forecasts to

18:08

anticipatory action and as you said we

18:11

can have these long range forecasts and

18:13

also chipping in on what Florence

18:15

mentioned about how you know weather is

18:18

quite chaotic so there's lots of

18:19

possibilities that can arise from a

18:21

given starting point so we have a lot of

18:23

uncertainty and you need to actually

18:26

generate a lot of different weather

18:27

forecasts to explore that AI allows you

18:30

to do that in a very lowcost way so you

18:33

can generate forecasts that explore the

18:37

uncertainty space in a very lowcost

18:39

manner so that you can actually properly

18:41

inform anticipatory action in these

18:44

areas we've talked a lot about weather

18:47

what we've not actually talked a lot

18:48

about is is climate change and and of

18:51

course there are climate deniers out

18:53

there Florence uh that we must uh

18:56

acknowledge I'm got to put a picture on

18:58

screen do you ever remember this uh this

19:01

was a a tornado that was coming at the

19:03

Florida Panhandle um and also was some

19:07

questions about whether it might go to

19:08

Alabama and and they got a Sharpie out

19:10

and they actually drew it on the end at

19:12

the Trump Administration which uh which

19:14

tells you that that uh you know we it

19:19

clearly is something that that people

19:21

try to play with when when we talk about

19:23

climate the way climate's changing what

19:26

weather is going to do but your your for

19:29

are so accurate Florence and presumably

19:31

with your digital twin earth You can

19:33

predict how climate is going to evolve

19:36

well into the future well exactly and we

19:39

use the same weather model to to do

19:42

climate models as well so they are just

19:44

a bit more complicated but it it's based

19:46

on the same sort of of modeling but also

19:49

we can document climate change and

19:51

that's what we're doing with the

19:53

Copernicus program from the EU going

19:56

back since you know from 1940 and really

19:59

depicting what the weather and climate

20:01

have been doing every hour from 1940 to

20:04

now we have this picture of the Earth

20:07

and we can then document how much the

20:09

temperature have increased how much the

20:11

frequency of storms have increased Etc

20:14

so it's predicting it but there is

20:16

already this reality we not we have

20:19

enough information to know what has had

20:22

happened already and then with these

20:24

models then we can do a digital twin of

20:27

of the climate as well and go forward in

20:29

the future with different scenarios of

20:32

course of what will happen in the

20:34

reduction of greenhouse gases because of

20:37

course it all depends how much we can

20:39

reduce the amount of carbon dioxide in

20:43

particular that we put in the

20:44

atmosphere and I've got a question for

20:46

all three of our distinguished scientist

20:48

only got a minute left so you're going

20:49

to have to make it quick go we keep

20:50

hearing that we're running out of data

20:52

and that this is a big problem for AI

20:54

but I wonder if that's actually true

20:56

particularly when it comes to weather

20:58

climate change and biodiversity loss we

21:01

have to fix obviously the climate change

21:03

and biodiversity problem do you feel

21:05

that there's a way for citizen

21:07

scientists to get back into action and

21:09

be submitting data to all of you

21:11

scientists so that you can help us fight

21:14

these bigger problems St Stephen pick

21:16

that up because we've just about 30

21:17

seconds left yeah it's a great shout and

21:21

as we talked about earlier we've got the

21:22

weather on the web there's another great

21:24

crowdsourcing initiative to look at um

21:28

in and early sighting of insects around

21:31

the UK which we've also connected with

21:33

climate change um here at the Met Office

21:36

along with many other partner

21:37

organizations so I think it's a great

21:38

shout for citizen science this one I

21:42

could talk plenty more as I always could

21:43

on this program every week we we never

21:45

get to the bottom of everything but

21:47

listen uh Florence Stephen Dr Nath

21:50

Stephanie thank you all very much uh for

21:51

your time really fascinating discussion

21:53

just a reminder we are putting all these

21:55

half hour programs on the BBC's Aid

21:57

decoded YouTube

21:59

site so you can find all our pass

22:00

programs there we'll do it again same

22:02

time next week thanks for watching