00:00 on September 11th 2023 weather
00:03 predictions in the northeast of the US
00:05 sounded like this all eyes are on
00:07 Hurricane Le the storm has strengthened
00:09 back to a category 3 we're expecting it
00:12 to make a northward turn over the next
00:14 few days by September 16th after being
00:17 downgraded storm Lee made landfall in
00:20 Nova Scotia Canada flooding roads
00:22 Downing trees and cutting out power for
00:25 tens of thousands of people along the
00:27 coast at least 5 days before hurricane
00:30 Lee struck land weather forecast had
00:32 roughly predicted its trajectory but
00:34 another forecast beat them 3 days before
00:37 weather stations an AI model created by
00:40 Google predicted the Cyclone's path the
00:43 AI revolution has reached meteorology
00:46 and it's at a time when we are
00:47 responding to extreme weather more than
00:49 ever we're about to find out if it'll
00:52 help us prepare by bringing the future
01:05 predicting future weather more than a
01:07 few hours out starts with creating a
01:09 snapshot of Earth's current atmosphere
01:11 scientists do that by collecting data
01:13 from sources Like Satellites and weather
01:15 stations and buoys located around the
01:17 world taking pictures of clouds and
01:19 measuring temperature and pressure and
01:21 wind speed and humidity all that
01:23 disperate data gets fed into computers
01:26 which generate a 3D grid of boxes that
01:28 represent the atmosphere both vertically
01:31 and horizontally computers then do a lot
01:33 of physics to determine how these
01:35 conditions interact with each other and
01:37 they produce a forecast I think any
01:39 forecast has like 150 trillion
01:42 calculations it's pretty amazing all
01:45 that math requires some of the world's
01:46 most powerful supercomputers the two big
01:49 ones are run by the European Center for
01:50 medium-range weather forecasts and the
01:52 National Weather Service in the US to
01:55 make a local weather forecast from this
01:57 Global model meteorologists zoom in and
01:59 ref find their own forecast with their
02:01 local expertise like if they live in a
02:04 hilly area or a flat area or near a lake
02:06 they'll adjust those models and do their
02:08 own professional interpretation based on
02:11 their area no matter what this initial
02:13 3D grid of atmosphere is never going to
02:16 exactly replicate reality there's too
02:18 many gaps in the data we can measure
02:20 that means forecasts get blurrier the
02:22 further out you go which is why the big
02:24 weather centers don't just generate one
02:26 forecast they tweak the initial data and
02:28 produce up to 50 forecasts it's called
02:31 Ensemble forecasting and it helps
02:33 meteorologists measure uncertainty if
02:35 all 50 forecasts look similar there's a
02:37 higher certainty in the prediction but
02:39 if there's a lot of variation there's
02:41 much less we got to kind of keep an eye
02:43 to the sky there's a potential up under
02:44 the storm in the works this one from 0%
02:46 chance to 40% at 2:00 70% at 8:00 and at
02:52 11:00 we had a tropical storm weather
02:55 centers only release their forecast
02:57 every 6 hours because that's all today's
02:59 computing power will allow but what if
03:01 that limmit didn't exist before we
03:03 explain that we'll hear from the sponsor
03:05 of this video this episode is presented
03:07 by Microsoft co-pilot for Microsoft 365
03:11 your AI assistant atwork co-pilot can
03:13 help you solve your most complex
03:15 problems at work going far beyond simple
03:17 questions and answers from getting up to
03:19 speed on a missed teams meeting in
03:21 seconds to helping you start a first
03:23 draft faster in word co-pilot for
03:26 Microsoft 365 gives everyone an AI
03:29 assistant netw work in their most
03:30 essential apps to unleash creativity
03:33 unlock productivity and uplevel skills
03:36 and it's all built on Microsoft's
03:37 comprehensive approach to security
03:40 privacy compliance and responsible AI
03:43 Microsoft does not influence the
03:44 editorial process of our videos but they
03:46 do help make videos like this possible
03:48 to learn more you can go to
03:51 microsoft.com foror now back to our
03:54 video in 2020 a group of researchers
03:56 published a data set called era 5 which
03:59 contained about 40 Years of the Earth's
04:00 hourly weather data that data set was
04:02 just primed for using AI cuz it was huge
04:06 the data is nice and smooth every
04:08 there's no missing values and it's free
04:11 just to step in like hey you have
04:13 terabytes of data I can learn how
04:14 weather moves AI models learned how
04:16 weather moves not through applying
04:18 trillions of physics equations to the
04:20 Earth's atmosphere but by being trained
04:22 on a5's enormous historical data set
04:25 researchers gave the models a snapshot
04:27 of weather conditions asked them to make
04:29 a prediction and then scored them on how
04:31 closely that prediction matched what
04:33 really happened after a while the models
04:35 eventually got really good at this by
04:37 2023 the tech companies Google Huawei
04:40 and Nvidia had developed models that
04:42 rivaled traditional forecasting on
04:44 variables like surface temperature
04:46 humidity and wind speed and on some
04:48 extreme weather events like the paths of
04:50 tropical Cyclones atmospheric rivers and
04:52 extreme temperatures these AI models
04:54 still rely on the same observation data
04:56 from the big weather centers the data
04:58 that creates that initial 3D grid
05:00 snapshot but they don't require anything
05:02 close to 6 hours to produce a prediction
05:05 huawei's pangu weather model for example
05:07 can produce a week-long forecast in 1.4
05:11 seconds which means that we spent over a
05:13 century figuring out the physics the
05:15 atmospheric science and the
05:16 computational skill to bring us our
05:18 modern day weather forecast and now
05:20 suddenly we have these AI models that
05:22 have come out of you know the past 2 3
05:24 years and they're getting the same skill
05:27 and now they run on a modest lap
05:30 top despite the impressive results from
05:32 these first AI models there's still lots
05:34 of work to do Google's graph cast
05:36 predicted hurricane Le's path faster
05:38 than traditional models but it didn't
05:41 prove it could predict A hurricane's
05:42 intensity which is a trickier
05:44 calculation to make these AI models are
05:47 incentivized to get as many correct
05:49 answers as they can through the scoring
05:50 system if you swing for it swing for the
05:52 fences right if it misses the model is
05:55 penalized very large it says no you
05:57 should never do that don't swing for the
05:58 fences cuz the error is going to be huge
06:00 but by prioritizing safer correct
06:02 answers to boosta model score it could
06:05 miss rare outlier weather events plus
06:07 they are learning from 40 Years of
06:09 history and historical weather has fewer
06:11 extreme events than we do today or will
06:13 have in the future due to climate change
06:16 but a big reason for optimism with these
06:18 AI models comes from their Ensemble
06:20 forecasting instead of the traditional
06:22 50 Ensemble forecasts they can predict a
06:25 thousand or more because they're freed
06:27 from Computing and time constraints
06:29 there's always going to be uncertainty
06:30 in a weather prediction but larger
06:32 ensembles will help us measure that
06:34 uncertainty better that's extremely
06:36 useful context to say you are a
06:38 emergency manager down in Florida who's
06:40 dealing with the very difficult decision
06:41 are are you going to you know order an
06:43 evacuation or not you want as much
06:45 information about the uncertainty as
06:47 possible large ensembles might also
06:49 catch a rare weather event that a 50
06:52 member Ensemble would miss or measure
06:54 the probability of weather events even
06:56 further into the future than our 10-day
06:58 forecast are we magically going to get a
07:00 crystal ball that lets us foresee
07:02 perfectly into the future probably not I
07:05 think especially on like subseasonal
07:06 time scales like multiple months out
07:09 we're going to be able to frame the
07:10 statistical question with a lot more
07:13 specificity and probably a much better
07:15 quantification of the uncertainty we do
07:17 have a new winter storm warning that's
07:18 been issued one thing we shouldn't
07:20 expect to change anytime soon is the
07:22 role of the meteorologist at least the
07:24 ones you see on TV if only because we
07:26 fundamentally have to communicate
07:28 uncertainty and we have to walk through
07:30 all the the various wh ifs and a human
07:31 is the best tool that we have today to
07:34 effectively communicate that and help
07:36 somebody else make a decision AI
07:37 forecasting models are still in an
07:39 experimental phase but the European
07:41 Center for medium-range weather
07:42 forecasting has started publishing AI
07:44 forecast alongside their traditional
07:46 ones for the public to compare when we
07:48 check the weather in the very near
07:50 future it might be powered by AI instead
07:52 of physics based models or a combination
07:54 of the two and if we get things right
07:56 we'll have a sharper view of the weather
07:58 events that we need to prepare for the
