WEBVTT 00:00:00.000 --> 00:00:02.020 Okay, let's just jump right into it. There's 00:00:02.020 --> 00:00:05.080 a brand new AI model entering the world of tropical 00:00:05.080 --> 00:00:08.199 cyclone forecasting that frankly could be absolutely 00:00:08.199 --> 00:00:10.980 huge for predicting these incredibly dangerous 00:00:10.980 --> 00:00:15.560 storms. But, and this is a big but, if it performs 00:00:15.560 --> 00:00:18.140 well in real time this season. That's the key 00:00:18.140 --> 00:00:20.480 test right there, absolutely. Welcome, everyone, 00:00:20.600 --> 00:00:23.800 to Meteorology Matters. This is where we really 00:00:23.800 --> 00:00:26.399 try to take an in -depth look at complex information, 00:00:26.600 --> 00:00:28.719 sort of unpack it, and help you get informed 00:00:28.719 --> 00:00:31.379 quickly and, well, thoroughly. Yep, cutting through 00:00:31.379 --> 00:00:34.219 the noise. Today, we're diving into something 00:00:34.219 --> 00:00:37.500 really fascinating and, honestly, often frustratingly 00:00:37.500 --> 00:00:40.969 difficult. the science of forecasting tropical 00:00:40.969 --> 00:00:43.450 cyclones. You might know them as hurricanes, 00:00:43.810 --> 00:00:46.729 typhoons, cyclones, depends where you are. Exactly. 00:00:46.950 --> 00:00:49.570 And we're exploring how artificial intelligence, 00:00:49.850 --> 00:00:53.369 AI, is now being applied in some, well, experimental 00:00:53.369 --> 00:00:56.240 ways, but ways that could... potentially revolutionize 00:00:56.240 --> 00:00:58.200 how we tackle this huge challenge. We're looking 00:00:58.200 --> 00:01:00.920 at some genuinely promising new approaches here, 00:01:01.060 --> 00:01:03.060 stuff that could really represent a leap forward 00:01:03.060 --> 00:01:05.519 in how accurate our predictions can be. Just 00:01:05.519 --> 00:01:07.459 take a second and think about the sheer power 00:01:07.459 --> 00:01:10.459 of these storms, the impact they have. Oh, it's 00:01:10.459 --> 00:01:12.680 immense, the danger, the destruction. And the 00:01:12.680 --> 00:01:15.840 cost. We were looking at some figures, an estimated 00:01:15.840 --> 00:01:20.480 $1 .4 trillion in economic losses globally just 00:01:20.480 --> 00:01:22.680 in the last 50 years. It's almost hard to wrap 00:01:22.680 --> 00:01:24.439 your head around a number like that. It really 00:01:24.439 --> 00:01:27.280 is. And behind that number, you've got countless 00:01:27.280 --> 00:01:30.019 lives disrupted, communities just devastated, 00:01:30.480 --> 00:01:33.019 long, hard roads to recovery. Yeah. And that 00:01:33.019 --> 00:01:35.980 really highlights why accurate warnings, why 00:01:35.980 --> 00:01:38.760 preparedness aren't just, you know, nice to haves. 00:01:39.000 --> 00:01:41.260 They're absolutely vital. Life or death, often. 00:01:41.400 --> 00:01:44.560 Totally. The earlier and more precisely we can 00:01:44.560 --> 00:01:46.680 figure out where these things are going, how 00:01:46.680 --> 00:01:49.489 strong they'll get. that directly translates 00:01:49.489 --> 00:01:52.810 into protecting people, securing property, saving 00:01:52.810 --> 00:01:55.129 lives. That's the whole point, right? That's 00:01:55.129 --> 00:01:56.890 the motivation behind all this forecasting work. 00:01:57.109 --> 00:01:59.230 And improving that forecasting, well, it's a 00:01:59.230 --> 00:02:02.150 serious scientific hurdle. Tropical cyclones 00:02:02.150 --> 00:02:03.909 aren't simple weather systems. You can just, 00:02:03.909 --> 00:02:06.170 you know, plug into a basic formula. Right, they're 00:02:06.170 --> 00:02:09.280 beasts. They really are. They're vast, complex, 00:02:09.280 --> 00:02:12.680 atmospheric engines. They're fuel. Primarily 00:02:12.680 --> 00:02:15.780 warm ocean water. Like giant heat pumps almost. 00:02:15.879 --> 00:02:19.120 That's a great analogy, actually. Giant, swirling 00:02:19.120 --> 00:02:21.680 heat pumps. They take heat and moisture from 00:02:21.680 --> 00:02:24.259 the sea surface, pull it way up into the atmosphere, 00:02:24.539 --> 00:02:27.560 and that drives this intense convection. Those 00:02:27.560 --> 00:02:29.699 huge thunderstorms spiraling around the core. 00:02:29.740 --> 00:02:32.979 And the scale is just massive. Huge. The strongest 00:02:32.979 --> 00:02:35.819 winds are near the center. Yeah. but dangerous 00:02:35.819 --> 00:02:38.360 conditions can stretch out for hundreds of miles. 00:02:38.719 --> 00:02:41.159 So what makes them so hard to forecast accurately? 00:02:41.379 --> 00:02:43.240 You said it's a hurdle. Well, a big part of it 00:02:43.240 --> 00:02:45.819 is their sensitivity, their incredibly finely 00:02:45.819 --> 00:02:48.240 tuned systems, their behavior, where they go, 00:02:48.319 --> 00:02:50.979 how strong they get. It's critically dependent 00:02:50.979 --> 00:02:53.560 on even really subtle differences in the atmosphere 00:02:53.560 --> 00:02:55.620 around them. Like what kind of differences? Oh, 00:02:55.659 --> 00:02:57.719 you know, a small shift in a steering current 00:02:57.719 --> 00:03:01.340 way up high, maybe miles away, a tiny change 00:03:01.340 --> 00:03:03.919 in how fast the ocean underneath is warming up, 00:03:03.919 --> 00:03:06.240 even just interacting with land if they get close. 00:03:06.360 --> 00:03:08.319 Any of these seemingly small things can cause 00:03:08.319 --> 00:03:10.479 a pretty dramatic change in the storm's path 00:03:10.479 --> 00:03:13.060 or, you know, how quickly it spins up or weakens. 00:03:13.300 --> 00:03:16.780 OK, so you have this massive, powerful, but also 00:03:16.780 --> 00:03:19.530 really sensitive system moving through and atmosphere 00:03:19.530 --> 00:03:22.030 that's constantly changing anyway. Exactly. That 00:03:22.030 --> 00:03:24.729 inherent complexity, that sensitivity, it's a 00:03:24.729 --> 00:03:27.250 big reason why the traditional forecasting models, 00:03:27.409 --> 00:03:29.389 the physics -based ones that are the workhorses 00:03:29.389 --> 00:03:31.990 today, have always faced this kind of fundamental 00:03:31.990 --> 00:03:34.610 trade -off. A trade -off in trying to predict 00:03:34.610 --> 00:03:37.650 everything about the storm. Yeah, exactly. Can 00:03:37.650 --> 00:03:41.189 you maybe break down what that sort of dual nature 00:03:41.189 --> 00:03:43.409 of forecasting means for these models? Sure. 00:03:43.810 --> 00:03:46.110 So these physics -based models, they're built 00:03:46.110 --> 00:03:48.949 on, well, fundamental physical laws equations 00:03:48.949 --> 00:03:52.009 right right that simulate how the atmosphere 00:03:52.009 --> 00:03:54.590 the ocean the land surface behave they create 00:03:54.590 --> 00:03:57.310 like a virtual earth basically and run time forward 00:03:57.310 --> 00:03:59.789 to see how weather evolves right but the problem 00:03:59.789 --> 00:04:02.569 is you can't model the entire planet done every 00:04:02.569 --> 00:04:05.789 tiny detail simultaneously not if you want the 00:04:05.789 --> 00:04:08.750 forecast you know now not next week the computing 00:04:08.750 --> 00:04:10.930 power just isn't there for that kind of real 00:04:10.930 --> 00:04:13.449 -time operational need even with today's supercomputers 00:04:13.449 --> 00:04:16.209 yeah You hit limit. So the model builders, they 00:04:16.209 --> 00:04:18.829 have to make choices. Big choices. Especially 00:04:18.829 --> 00:04:22.089 about resolution. How detailed the model's view 00:04:22.089 --> 00:04:25.050 is. Okay. So predicting the cyclone's track, 00:04:25.230 --> 00:04:26.629 right, where it's generally going to go over 00:04:26.629 --> 00:04:28.709 several days. That's mostly determined by the 00:04:28.709 --> 00:04:31.850 big picture stuff. The large scale atmospheric 00:04:31.850 --> 00:04:35.069 flow, these huge currents of air high up that 00:04:35.069 --> 00:04:37.250 basically steer the storm along. Like rivers 00:04:37.250 --> 00:04:40.110 in the sky pushing it. Exactly. So to capture 00:04:40.110 --> 00:04:42.889 those global steering currents properly, you 00:04:42.889 --> 00:04:45.389 need a model that covers a huge area, ideally, 00:04:45.709 --> 00:04:48.610 the whole globe. A global model. Right. But because 00:04:48.610 --> 00:04:50.269 it's global, it has to have lower resolution. 00:04:50.430 --> 00:04:52.949 It sees the atmosphere in bigger chunks, maybe 00:04:52.949 --> 00:04:55.850 grid boxes tens of kilometers wide. It gets the 00:04:55.850 --> 00:04:58.589 forest view, the overall flow, which is great 00:04:58.589 --> 00:05:00.769 for predicting the track. Makes sense. Global 00:05:00.769 --> 00:05:02.850 view for global steering. But then there's the 00:05:02.850 --> 00:05:04.649 intensity. Is it going to stay a tropical storm 00:05:04.649 --> 00:05:07.389 or blow up into a major category four or five 00:05:07.389 --> 00:05:10.379 hurricane? The million dollar question. Often. 00:05:10.600 --> 00:05:13.819 Right. And that depends on incredibly complex, 00:05:14.220 --> 00:05:16.699 really small -scale things happening inside and 00:05:16.699 --> 00:05:19.180 right around the storm's core. The eye, the eye 00:05:19.180 --> 00:05:21.800 wall. We're talking about how air is rising and 00:05:21.800 --> 00:05:24.759 sinking in these intense thunderstorm bursts, 00:05:25.480 --> 00:05:28.319 how moisture is cycling through the turbulent 00:05:28.319 --> 00:05:30.980 winds mixing near the surface, how the storm 00:05:30.980 --> 00:05:33.040 is interacting with the ocean right below it, 00:05:33.199 --> 00:05:37.060 even the friction from waves. Tiny details, relatively 00:05:37.060 --> 00:05:39.769 speaking. Stuff happening on a much finer scale 00:05:39.769 --> 00:05:42.389 than those big steering currents. Exactly. And 00:05:42.389 --> 00:05:44.509 to simulate those intricate details, you need 00:05:44.509 --> 00:05:47.529 models with really high resolution. Models focused 00:05:47.529 --> 00:05:50.370 specifically on the storm and its immediate surroundings. 00:05:50.730 --> 00:05:53.370 So, regional models. Typically, yeah. Regional 00:05:53.370 --> 00:05:55.750 models. They cover a smaller area, but their 00:05:55.750 --> 00:05:57.870 grid boxes might only be a few kilometers wide. 00:05:57.910 --> 00:06:00.509 They can see the trees, maybe even the leaves, 00:06:01.009 --> 00:06:02.990 of the storm's internal structure, which you 00:06:02.990 --> 00:06:05.970 need for intensity. Okay, so... global lower 00:06:05.970 --> 00:06:08.509 resolution for the big picture track and regional 00:06:08.509 --> 00:06:11.230 high resolution for the small scale intensity 00:06:11.230 --> 00:06:13.069 details. Yeah. That's the trade off you mentioned. 00:06:13.230 --> 00:06:15.629 That's the classic trade off. You generally can't 00:06:15.629 --> 00:06:19.209 have one physics based model that does both optimally, 00:06:19.569 --> 00:06:21.930 especially not fast enough for operational forecasting. 00:06:22.269 --> 00:06:24.129 Why not? I mean, why can't you just run a super 00:06:24.129 --> 00:06:28.399 high res global model? Just too slow. too computationally 00:06:28.399 --> 00:06:31.319 expensive. To run fast enough to be useful in 00:06:31.319 --> 00:06:34.319 real time, physics models have to make approximations. 00:06:34.779 --> 00:06:37.860 You can build that global model, but its resolution 00:06:37.860 --> 00:06:40.980 will be too coarse, too fuzzy to really capture 00:06:40.980 --> 00:06:43.680 those fine details driving intensity accurately. 00:06:43.699 --> 00:06:46.120 All right, it misses the leaves. And you can 00:06:46.120 --> 00:06:48.319 build that high resolution regional model focusing 00:06:48.319 --> 00:06:52.230 on the storm. But two problems. One, it still 00:06:52.230 --> 00:06:54.790 needs information from a global model at its 00:06:54.790 --> 00:06:57.329 edges to know how the larger environment is pushing 00:06:57.329 --> 00:07:00.050 it around, the boundary conditions. Okay, it 00:07:00.050 --> 00:07:02.910 needs the big picture context. And two, its high 00:07:02.910 --> 00:07:06.050 resolution still makes it too computationally 00:07:06.050 --> 00:07:09.490 demanding to run for the whole globe. So what 00:07:09.490 --> 00:07:11.430 operational centers usually do is run a whole 00:07:11.430 --> 00:07:13.149 suite of different models. Right. You see those 00:07:13.149 --> 00:07:15.290 spaghetti plots sometimes. Exactly. That often 00:07:15.290 --> 00:07:17.269 comes from ensemble systems where they run the 00:07:17.269 --> 00:07:20.250 same model maybe 20, 30, 50 times with slightly 00:07:20.250 --> 00:07:21.850 different starting conditions to get a range 00:07:21.850 --> 00:07:23.829 of possibilities. Plus they look at different 00:07:23.829 --> 00:07:26.230 global models, different regional models. Trying 00:07:26.230 --> 00:07:28.670 to piece together the best possible forecast 00:07:28.670 --> 00:07:31.569 from all these different tools. Precisely. But 00:07:31.569 --> 00:07:34.709 it's been this persistent challenge getting truly 00:07:34.709 --> 00:07:37.329 state -of -the -art accuracy for both track and 00:07:37.329 --> 00:07:39.970 intensity at the same time from a single physics 00:07:39.970 --> 00:07:42.290 -based system because of those computational 00:07:42.290 --> 00:07:44.370 limits and the different scales of physics involved. 00:07:44.509 --> 00:07:47.470 Which brings us back to why any improvement matters 00:07:47.470 --> 00:07:50.790 so much. Absolutely. When a storm's bearing down, 00:07:50.990 --> 00:07:53.009 knowing if it's going to hit 50 miles further 00:07:53.009 --> 00:07:55.250 north or if it's suddenly going to ramp up in 00:07:55.250 --> 00:07:58.370 strength just before landfall. That makes a world 00:07:58.370 --> 00:08:01.470 of difference. For evacuations, for staging resources, 00:08:01.769 --> 00:08:03.689 for just telling people how serious the risk 00:08:03.689 --> 00:08:06.509 is. It really does. Better predictions directly 00:08:06.509 --> 00:08:09.610 save lives and cut down on that staggering economic 00:08:09.610 --> 00:08:12.009 damage we talked about. Every kilometer closer 00:08:12.009 --> 00:08:14.269 on the track, every knot more accurate on the 00:08:14.269 --> 00:08:16.410 intensity. It translates into better decisions 00:08:16.410 --> 00:08:19.050 on the ground, which is why this new development 00:08:19.050 --> 00:08:20.810 we're talking about is getting so much buzz. 00:08:21.250 --> 00:08:23.189 It's a fundamentally different way of tackling 00:08:23.189 --> 00:08:25.139 the problem. All right, let's dig into that, 00:08:25.339 --> 00:08:27.920 this different approach. Tell us about this new 00:08:27.920 --> 00:08:31.319 experimental AI -based tropical cyclone model. 00:08:31.500 --> 00:08:33.700 What's the story here? OK, so this comes out 00:08:33.700 --> 00:08:36.019 of a collaboration between Google DeepMind and 00:08:36.019 --> 00:08:39.320 Google Research. And the key difference is, unlike 00:08:39.320 --> 00:08:42.139 those tradition models trying to explicitly calculate 00:08:42.139 --> 00:08:45.700 physics equations, this AI model learns directly 00:08:45.700 --> 00:08:48.379 from massive amounts of historical weather data. 00:08:48.840 --> 00:08:50.580 The source material we looked at says it's based 00:08:50.580 --> 00:08:52.879 on something called stochastic neural networks. 00:08:53.100 --> 00:08:56.100 Stochastic neural networks. OK, that sounds technical. 00:08:56.559 --> 00:08:58.679 Can you give us the simpler version? What does 00:08:58.679 --> 00:09:01.860 that actually mean for forecasting? Sure. Think 00:09:01.860 --> 00:09:04.879 of neural networks at a basic level as really 00:09:04.879 --> 00:09:07.980 powerful pattern recognition machines. You train 00:09:07.980 --> 00:09:10.740 them on enormous data sets, and they learn complex 00:09:10.740 --> 00:09:12.779 relationships and patterns within that data. 00:09:13.080 --> 00:09:15.220 Without being explicitly told the physics equations? 00:09:15.500 --> 00:09:17.460 Kind of, yeah. They learn the correlations that 00:09:17.460 --> 00:09:19.539 cause and effect relationships as reflected in 00:09:19.539 --> 00:09:22.820 the data itself. Now, the stochastic part, that 00:09:22.820 --> 00:09:26.279 basically means it involves probability or randomness. 00:09:26.460 --> 00:09:29.259 So in this context, it means the model is designed 00:09:29.259 --> 00:09:32.529 to handle the inherent Uncertainty in weather. 00:09:33.409 --> 00:09:36.210 Weather isn't perfectly predictable, right? So 00:09:36.210 --> 00:09:38.570 a stochastic model doesn't just spit out one 00:09:38.570 --> 00:09:41.529 single answer. It can generate a range of possible 00:09:41.529 --> 00:09:43.909 outcomes. Like those ensemble models you mentioned. 00:09:44.169 --> 00:09:46.570 Exactly. It reflects the probabilistic nature 00:09:46.570 --> 00:09:49.389 of the atmosphere. So it's a sophisticated type 00:09:49.389 --> 00:09:51.929 of AI that's particularly well suited for trying 00:09:51.929 --> 00:09:54.970 to model chaotic, dynamic systems like weather. 00:09:55.429 --> 00:09:58.049 It loans the patterns and acknowledges the uncertainty. 00:09:58.350 --> 00:10:00.980 Got it. Learning patterns from data. Handling 00:10:00.980 --> 00:10:03.659 uncertainty giving multiple possibilities. So 00:10:03.659 --> 00:10:06.779 how does this AI approach? Claim to get around 00:10:06.779 --> 00:10:09.899 that classic trade -off being good at both track 00:10:09.899 --> 00:10:12.200 and intensity in one system Well, this is really 00:10:12.200 --> 00:10:13.779 the core innovation they're highlighting and 00:10:13.779 --> 00:10:16.320 it's kind of the aha moment This experimental 00:10:16.320 --> 00:10:19.519 AI model is designed as a single unified system 00:10:19.519 --> 00:10:21.820 one model for both. That's the goal. Yeah aiming 00:10:21.820 --> 00:10:23.860 for state -of -the -art accuracy for both track 00:10:23.860 --> 00:10:26.639 and intensity predictions simultaneously. Instead 00:10:26.639 --> 00:10:28.940 of being limited by the computational cost of 00:10:28.940 --> 00:10:30.820 simulating high -resolution physics everywhere, 00:10:31.000 --> 00:10:34.320 like the traditional models. Right. This AI model, 00:10:34.700 --> 00:10:36.879 through its training on all that data, seems 00:10:36.879 --> 00:10:39.240 to be capable of learning the really complex 00:10:39.240 --> 00:10:41.679 interplay between the large -scale atmospheric 00:10:41.679 --> 00:10:44.059 patterns that steer the storm, the forest view, 00:10:44.200 --> 00:10:46.919 and the fine -scale internal dynamics that control 00:10:46.919 --> 00:10:50.659 its intensity. the leaves you. Exactly. It appears 00:10:50.659 --> 00:10:53.539 to learn these connections directly from observations 00:10:53.539 --> 00:10:56.340 across all those different scales, rather than 00:10:56.340 --> 00:10:58.940 having to explicitly simulate the physics with 00:10:58.940 --> 00:11:00.759 different resolutions for different aspects. 00:11:00.960 --> 00:11:02.840 It's just a different way of modeling the problem. 00:11:02.960 --> 00:11:05.080 That does sound like a potential way around those 00:11:05.080 --> 00:11:07.759 computational bottlenecks, learning the whole 00:11:07.759 --> 00:11:10.799 complex dance from the data itself. That's the 00:11:10.799 --> 00:11:13.259 idea, potentially bypassing some of those limitations. 00:11:13.580 --> 00:11:16.899 So what specific things can this AI model predict? 00:11:17.129 --> 00:11:20.149 about a hurricane or typhoon. The source material 00:11:20.149 --> 00:11:22.230 lists the key things forecasters need to know. 00:11:22.409 --> 00:11:24.610 It can predict the storm's formation, basically, 00:11:25.090 --> 00:11:26.950 flagging when and where a weather disturbance 00:11:26.950 --> 00:11:29.929 is likely to organize into a named tropical cyclone. 00:11:30.370 --> 00:11:33.350 That's useful. Early warning. Very. Then, of 00:11:33.350 --> 00:11:35.549 course, the storms track its likely path, its 00:11:35.549 --> 00:11:37.450 intensity, how strong the winds will be over 00:11:37.450 --> 00:11:41.929 time. But also, crucially, it predicts the storm's 00:11:41.929 --> 00:11:44.350 size and its shape. Size and shape. Why are those 00:11:44.350 --> 00:11:47.110 important? Well, size tells you how large an 00:11:47.110 --> 00:11:49.190 area is going to be affected by dangerous winds, 00:11:49.529 --> 00:11:52.509 storm surge, heavy rain. A huge storm affects 00:11:52.509 --> 00:11:55.330 a much wider zone than a small, compact one, 00:11:55.450 --> 00:11:58.389 even if their peak winds are the same. OK. And 00:11:58.389 --> 00:12:00.889 shape influences how those hazards are distributed. 00:12:01.350 --> 00:12:03.809 Is the strongest wind quadrant going to be on 00:12:03.809 --> 00:12:06.110 the right side, the left side? No. That matters 00:12:06.110 --> 00:12:09.149 for pinpointing the greatest impact areas. So 00:12:09.149 --> 00:12:11.629 predicting those five things, formation, track, 00:12:11.909 --> 00:12:14.710 intensity, size, and shape, gives a pretty comprehensive 00:12:14.710 --> 00:12:17.429 picture for forecasters. Cover the bases. And 00:12:17.429 --> 00:12:18.889 you mentioned earlier, it doesn't just give one 00:12:18.889 --> 00:12:20.330 prediction. You said something about multiple 00:12:20.330 --> 00:12:22.649 outcomes. Right. That's another really key feature 00:12:22.649 --> 00:12:25.070 they mentioned. Instead of just outputting one 00:12:25.070 --> 00:12:27.350 single forecast track and intensity profile, 00:12:27.470 --> 00:12:29.830 like a traditional deterministic model run might, 00:12:30.529 --> 00:12:33.330 this AI model can generate 50 possible scenarios 00:12:33.330 --> 00:12:36.330 for a single cyclone at each forecast time step. 00:12:36.850 --> 00:12:39.629 50. Wow. So what does that look like? Imagine 00:12:39.629 --> 00:12:42.350 looking at a forecast map instead of just one 00:12:42.350 --> 00:12:45.009 line showing the most likely track. or one cone 00:12:45.009 --> 00:12:48.149 of uncertainty, you might see 50 slightly different 00:12:48.149 --> 00:12:51.549 plausible tracks fanning out and maybe 50 different 00:12:51.549 --> 00:12:54.210 intensity forecasts associated with them. That 00:12:54.210 --> 00:12:57.470 sounds complex, but also really useful. It's 00:12:57.470 --> 00:13:00.269 incredibly valuable because, like we said, weather 00:13:00.269 --> 00:13:02.889 is uncertain. A single forecast line, no matter 00:13:02.889 --> 00:13:05.450 how good the model, kind of hides that uncertainty. 00:13:05.690 --> 00:13:08.190 Seeing 50 different possibilities, the spread 00:13:08.190 --> 00:13:10.850 in the tracks, the range of potential intensities 00:13:10.850 --> 00:13:13.590 gives forecasters and emergency managers a much 00:13:13.590 --> 00:13:16.190 better handle on the risk. They can see not just 00:13:16.190 --> 00:13:19.110 the most likely outcome, but also the plausible 00:13:19.110 --> 00:13:21.370 worst -case scenarios within that spread. Exactly. 00:13:21.470 --> 00:13:24.210 And maybe even some less likely, but still possible 00:13:24.210 --> 00:13:26.629 outlier scenarios. It allows for much more nuanced 00:13:26.629 --> 00:13:29.370 risk assessment and planning. It's just a richer, 00:13:29.549 --> 00:13:32.269 more realistic picture than a single prediction. 00:13:32.629 --> 00:13:35.120 50 scenarios. Definitely more information to 00:13:35.120 --> 00:13:37.519 work with. And how far ahead does it try to predict? 00:13:37.799 --> 00:13:39.820 What's the forecast horizon? The source says 00:13:39.820 --> 00:13:42.659 it provides predictions up to 15 days out. 15 00:13:42.659 --> 00:13:45.220 days? That's really long range for hurricane 00:13:45.220 --> 00:13:48.600 forecasting. It is. And, you know, accuracy naturally 00:13:48.600 --> 00:13:50.740 decreases the further out you go with any model. 00:13:51.340 --> 00:13:54.419 But even having a probabilistic glimpse of potential 00:13:54.419 --> 00:13:57.240 development or movement two weeks ahead... That 00:13:57.240 --> 00:14:00.360 can be incredibly useful for very early large 00:14:00.360 --> 00:14:02.519 -scale planning, thinking about positioning major 00:14:02.519 --> 00:14:05.019 assets, long lead time preparedness measures, 00:14:05.080 --> 00:14:07.659 that kind of thing. Okay, so the potential capabilities 00:14:07.659 --> 00:14:11.259 sound impressive. Unified model, track, intensity, 00:14:11.659 --> 00:14:14.519 size, shape, probabilistic forecast out to 15 00:14:14.519 --> 00:14:17.080 days. But how did they actually build this? What 00:14:17.080 --> 00:14:19.399 did it learn from? What's the data diet, as you 00:14:19.399 --> 00:14:21.679 called it? Right, because the AI is only as good 00:14:21.679 --> 00:14:24.039 as the data it learns from. And the source explains 00:14:24.039 --> 00:14:26.500 they use two main types of data, and critically, 00:14:26.700 --> 00:14:29.070 they use them together. OK, what were they? First, 00:14:29.169 --> 00:14:31.850 it was trained on a massive atmospheric reanalysis 00:14:31.850 --> 00:14:34.129 data set. Reanalysis. Yeah. What's that exactly? 00:14:34.409 --> 00:14:37.529 Reanalysis data sets are amazing resources. They 00:14:37.529 --> 00:14:39.669 essentially reconstruct the state of the Earth's 00:14:39.669 --> 00:14:41.610 atmosphere over many, many years, sometimes going 00:14:41.610 --> 00:14:44.490 back decades. They do this by taking millions 00:14:44.490 --> 00:14:47.269 and millions of historical observations. From 00:14:47.269 --> 00:14:49.909 satellites, weather stations, planes, ships. 00:14:50.009 --> 00:14:52.929 All of that, yeah. Balloons, buoys, everything. 00:14:53.149 --> 00:14:56.629 and they combine or assimilate all those observations 00:14:56.629 --> 00:14:59.690 into a sophisticated physics -based weather model. 00:15:00.450 --> 00:15:03.610 The result is a consistent, long -term, hour 00:15:03.610 --> 00:15:06.450 -by -hour history of the weather across the entire 00:15:06.450 --> 00:15:09.110 globe. So it's like the best possible reconstruction 00:15:09.110 --> 00:15:12.460 of past weather. Pretty much. And by training 00:15:12.460 --> 00:15:15.200 on this massive data set, the AI model learned 00:15:15.200 --> 00:15:17.220 about the general behavior of the atmosphere, 00:15:17.799 --> 00:15:20.220 global patterns, seasonal cycles, large -scale 00:15:20.220 --> 00:15:22.879 features like jet streams, those big steering 00:15:22.879 --> 00:15:25.669 currents we talked about. This provides the essential 00:15:25.669 --> 00:15:28.149 global context for predicting where a storm might 00:15:28.149 --> 00:15:30.070 go. Oh yeah, so that's the big picture context. 00:15:30.149 --> 00:15:31.909 What was the second type of data? The second 00:15:31.909 --> 00:15:34.470 was a specialized database focused purely on 00:15:34.470 --> 00:15:36.669 tropical cyclones. This contained the detailed 00:15:36.669 --> 00:15:38.769 historical records, the verified tracks, the 00:15:38.769 --> 00:15:41.230 measured intensities, the estimated size, the 00:15:41.230 --> 00:15:44.250 wind radii for nearly 5 ,000 actual tropical 00:15:44.250 --> 00:15:46.629 cyclones observed over the past 45 years or so. 00:15:46.669 --> 00:15:50.539 Wow, 5 ,000 storms? Yeah. So by studying all 00:15:50.539 --> 00:15:52.980 those past storms, the model learned the specific 00:15:52.980 --> 00:15:55.340 characteristics and behaviors of tropical cyclones 00:15:55.340 --> 00:15:58.600 themselves, how they tend to form, how they intensify 00:15:58.600 --> 00:16:01.179 or weaken under different conditions, how their 00:16:01.179 --> 00:16:03.440 structure, their size, and wind fields typically 00:16:03.440 --> 00:16:06.820 evolve. This historical storm data is key for 00:16:06.820 --> 00:16:08.899 getting the intensity and the storm structure 00:16:08.899 --> 00:16:12.259 predictions right. So it learned the global weather 00:16:12.259 --> 00:16:15.639 patterns and the specific details of thousands 00:16:15.639 --> 00:16:18.179 of real storms within those patterns. Exactly. 00:16:18.190 --> 00:16:20.669 And the source really emphasizes this point. 00:16:21.070 --> 00:16:23.070 Training the AI model on these two different 00:16:23.070 --> 00:16:25.480 types of data together was crucial. It didn't 00:16:25.480 --> 00:16:27.320 just learn about steering currents in isolation 00:16:27.320 --> 00:16:30.480 or past storm behavior in isolation. It learned 00:16:30.480 --> 00:16:33.580 how past storms actually behaved within the specific 00:16:33.580 --> 00:16:35.940 global atmospheric environments they moved through. 00:16:36.460 --> 00:16:39.000 That combination was key to improving its prediction 00:16:39.000 --> 00:16:41.279 capabilities compared to using just one data 00:16:41.279 --> 00:16:43.399 set alone. OK, that makes sense. Learning the 00:16:43.399 --> 00:16:45.899 context and the specific events together. So 00:16:45.899 --> 00:16:47.620 how well did this training actually work? Did 00:16:47.620 --> 00:16:49.600 they test it? How did it stack up against the 00:16:49.600 --> 00:16:53.279 models already in use? Ah yes, the all -important 00:16:53.279 --> 00:16:55.860 testing and comparison phase. This is where we 00:16:55.860 --> 00:16:58.159 see the concrete numbers suggesting its potential. 00:16:58.639 --> 00:17:01.559 The source details how they compared their experimental 00:17:01.559 --> 00:17:05.019 AI model against some of the leading operational 00:17:05.019 --> 00:17:07.619 physics -based models used by forecast centers 00:17:07.619 --> 00:17:10.019 around the world. Who did they compare it against? 00:17:10.160 --> 00:17:12.720 OK, so for track accuracy, they went head to 00:17:12.720 --> 00:17:16.079 head with ENS. ENS. That's the big European model 00:17:16.079 --> 00:17:18.819 ensemble. That's the one. The Global Ensemble 00:17:18.819 --> 00:17:21.140 System from the European Center for Medium Range 00:17:21.140 --> 00:17:24.880 Weather Forecasts, ECMWF. It's widely regarded 00:17:24.880 --> 00:17:27.940 as one of the best, if not the best, global weather 00:17:27.940 --> 00:17:30.099 prediction systems in the world, especially for 00:17:30.099 --> 00:17:32.500 track forecasting. So it's off benchmark. Absolutely. 00:17:32.980 --> 00:17:35.339 And they tested the AI model's track predictions 00:17:35.339 --> 00:17:38.059 against the actual verified paths of hurricanes. 00:17:38.220 --> 00:17:41.299 observed by NOAA, the National Oceanic and Atmospheric 00:17:41.299 --> 00:17:44.339 Administration, during the 2023 and 2024 hurricane 00:17:44.339 --> 00:17:47.180 seasons, specifically focusing on the North Atlantic 00:17:47.180 --> 00:17:49.900 and East Pacific basins. Okay, real -world test 00:17:49.900 --> 00:17:52.079 cases. What did they find? The results reported 00:17:52.079 --> 00:17:55.980 are... well... Pretty striking. On average, they 00:17:55.980 --> 00:17:58.420 found that the AI model's five -day track prediction 00:17:58.420 --> 00:18:01.779 was 140 kilometers closer to the storm's true 00:18:01.779 --> 00:18:04.200 observed location than the five -day prediction 00:18:04.200 --> 00:18:07.500 from the benchmark ENS model. 140 kilometers 00:18:07.500 --> 00:18:10.480 closer. At five days out, that sounds like a 00:18:10.480 --> 00:18:12.660 really significant improvement. Is it? It is 00:18:12.660 --> 00:18:14.710 significant, yeah. But the source gives even 00:18:14.710 --> 00:18:16.630 better context to understand how significant. 00:18:16.910 --> 00:18:19.630 They point out that this 140 kilometer reduction 00:18:19.630 --> 00:18:22.589 in the average five -day track error, well, that 00:18:22.589 --> 00:18:24.930 level of accuracy is comparable to what the ENS 00:18:24.930 --> 00:18:27.910 model typically achieves at only 3 .5 days out. 00:18:28.029 --> 00:18:31.430 Wait, so... The AI model at 5 days was as accurate 00:18:31.430 --> 00:18:34.289 as the leading global model is at 3 .5 days. 00:18:34.490 --> 00:18:36.750 That's a comparison they draw. Essentially, it 00:18:36.750 --> 00:18:38.630 suggests the AI model demonstrated something 00:18:38.630 --> 00:18:41.690 like a 1 .5 -day improvement in track forecast 00:18:41.690 --> 00:18:43.869 lead time accuracy. A day and a half better lead 00:18:43.869 --> 00:18:46.049 time for the same accuracy level. Wow. And the 00:18:46.049 --> 00:18:47.849 source puts that into perspective, too. They 00:18:47.849 --> 00:18:50.289 note that achieving that kind of 1 .5 -day improvement 00:18:50.289 --> 00:18:53.009 in track forecasting skill, historically, through 00:18:53.009 --> 00:18:54.910 gradual refinements of physics -based models, 00:18:55.230 --> 00:18:57.509 has typically taken over a decade. A decade's 00:18:57.509 --> 00:19:00.440 worth of progress? potentially achieve much faster 00:19:00.440 --> 00:19:03.099 with this AI approach. That's the implication 00:19:03.099 --> 00:19:05.799 from these internal tests. Seeing that magnitude 00:19:05.799 --> 00:19:08.599 of improvement, basically jumping forward maybe 00:19:08.599 --> 00:19:11.240 18 months worth of traditional progress in track 00:19:11.240 --> 00:19:14.339 accuracy from a totally new modeling paradigm. 00:19:14.720 --> 00:19:17.140 That's a really eye -catching metric. It underscores 00:19:17.140 --> 00:19:19.980 the potential scale of this advancement if it 00:19:19.980 --> 00:19:22.880 holds up. That really drives it home. Okay, so 00:19:22.880 --> 00:19:25.599 a potentially massive leap in track forecasting. 00:19:26.359 --> 00:19:28.700 What about the other half of the problem intensity? 00:19:28.920 --> 00:19:31.180 How did it do there? Right, the intensity forecast. 00:19:31.579 --> 00:19:34.359 For that comparison, they benchmarked their AI 00:19:34.359 --> 00:19:37.640 model against HEFS. HEFS. That's the Hurricane 00:19:37.640 --> 00:19:40.180 Analysis and Forecast System. Correct. That's 00:19:40.180 --> 00:19:42.500 a leading, regional, high -resolution, physics 00:19:42.500 --> 00:19:44.500 -based model developed specifically by NOAA, 00:19:44.660 --> 00:19:47.079 the National Oceanic and Atmospheric Administration. 00:19:47.279 --> 00:19:50.079 It's designed precisely to capture those complex, 00:19:50.240 --> 00:19:52.059 small -scale storm processes we talked about 00:19:52.059 --> 00:19:54.680 that govern intensity changes. So it's a specialist 00:19:54.680 --> 00:19:57.180 model for intensity. OK, so comparing the AI 00:19:57.180 --> 00:19:58.940 against the specialist intensity model. What 00:19:58.940 --> 00:20:01.519 was the outcome? According to the internal testing 00:20:01.519 --> 00:20:04.519 results in source, the experimental AI model 00:20:04.519 --> 00:20:07.519 actually outperformed the average intensity error 00:20:07.519 --> 00:20:10.579 of the HAFS model. It beat the specialist intensity 00:20:10.579 --> 00:20:13.079 model. In terms of average intensity error in 00:20:13.079 --> 00:20:16.019 their tests, yes. So the findings were twofold. 00:20:16.720 --> 00:20:19.359 A significant improvement in track compared to 00:20:19.359 --> 00:20:22.940 a top global model and better intensity accuracy 00:20:22.940 --> 00:20:26.000 compared to a leading regional high resolution 00:20:26.000 --> 00:20:28.279 model specifically built for that purpose. And 00:20:28.279 --> 00:20:31.430 all from that single unified AI system. That's 00:20:31.430 --> 00:20:34.109 the remarkable claim. It seems to suggest that, 00:20:34.109 --> 00:20:36.549 at least in these evaluations, the AI approach 00:20:36.549 --> 00:20:38.549 was successful in learning to handle both the 00:20:38.549 --> 00:20:41.289 large -scale steering factors for track and the 00:20:41.289 --> 00:20:44.390 fine -scale internal physics for intensity simultaneously, 00:20:44.869 --> 00:20:47.309 overcoming that traditional tradeoff. That is 00:20:47.309 --> 00:20:49.569 seriously impressive if it translates to consistent 00:20:49.569 --> 00:20:52.089 real -world performance. Exactly. The source 00:20:52.089 --> 00:20:54.410 also mentions briefly that preliminary tests 00:20:54.410 --> 00:20:57.009 for predicting the storm's size and wind radii 00:20:57.009 --> 00:20:58.990 showed results that were comparable with the 00:20:58.990 --> 00:21:01.970 existing physics -based models. Promising across 00:21:01.970 --> 00:21:04.369 the board basically. The initial internal results 00:21:04.369 --> 00:21:07.250 certainly paint that picture. They also note, 00:21:07.809 --> 00:21:10.410 importantly for the scientific process, that 00:21:10.410 --> 00:21:13.250 visualizations comparing these track and intensity 00:21:13.250 --> 00:21:17.049 errors are available for other experts to scrutinize. 00:21:17.509 --> 00:21:20.190 Transparency is key here. Absolutely. Okay so 00:21:20.190 --> 00:21:23.569 the internal tests look fantastic on paper, but 00:21:23.569 --> 00:21:26.769 as you said earlier, the real world is the ultimate 00:21:26.769 --> 00:21:29.859 proving ground. and getting buy -in and validation 00:21:29.859 --> 00:21:32.099 from the operational forecasters, the people 00:21:32.099 --> 00:21:34.359 who actually have to use this stuff to make critical 00:21:34.359 --> 00:21:37.460 decisions, that's essential. Couldn't agree more. 00:21:37.720 --> 00:21:39.920 You can have the best model in the lab. But if 00:21:39.920 --> 00:21:41.900 the operational community doesn't trust it or 00:21:41.900 --> 00:21:44.160 can't effectively use it, it doesn't help anyone 00:21:44.160 --> 00:21:46.859 when a storm is actually threatening. That brings 00:21:46.859 --> 00:21:48.859 us to the partnerships they highlighted. Right. 00:21:48.920 --> 00:21:51.099 Who are they working with to validate this? The 00:21:51.099 --> 00:21:53.339 source emphasizes several key collaborations. 00:21:53.700 --> 00:21:55.359 Probably the most prominent one mentioned is 00:21:55.359 --> 00:21:57.319 with the U .S. National Hurricane Center, the 00:21:57.319 --> 00:22:00.180 NHG. OK. The folks who issue the official forecast 00:22:00.180 --> 00:22:02.809 for the Atlantic and East Pacific. Exactly. They 00:22:02.809 --> 00:22:05.009 are the operational experts on the front lines 00:22:05.009 --> 00:22:08.029 for those regions. The source says the NHC has 00:22:08.029 --> 00:22:10.470 been scientifically validating the AI approach 00:22:10.470 --> 00:22:13.509 and its outputs. What does scientifically validating 00:22:13.509 --> 00:22:16.809 mean in practice? It means their expert meteorologists 00:22:16.809 --> 00:22:19.769 are looking closely at how the AI model performs, 00:22:20.269 --> 00:22:22.329 comparing its predictions against what actually 00:22:22.329 --> 00:22:25.089 happens with real storms, assessing its strengths, 00:22:25.250 --> 00:22:27.930 its weaknesses, its biases under different scenarios. 00:22:28.029 --> 00:22:30.529 And are they just looking at past data or? No, 00:22:30.529 --> 00:22:33.460 this is the really The source states that NHC 00:22:33.460 --> 00:22:36.140 expert forecasters are now seeing live predictions 00:22:36.140 --> 00:22:39.940 from these experimental AI models. Live. As in 00:22:39.940 --> 00:22:42.599 during this current hurricane season. Yes. They're 00:22:42.599 --> 00:22:44.900 seeing the AI output streams on their internal 00:22:44.900 --> 00:22:47.819 forecasting systems right alongside all the traditional 00:22:47.819 --> 00:22:50.799 physics -based model guidance and the real -time 00:22:50.799 --> 00:22:52.859 satellite and aircraft observations they use 00:22:52.859 --> 00:22:55.809 every day. Wow. So, the human experts making 00:22:55.809 --> 00:22:58.769 the official forecast decisions at the NHC are 00:22:58.769 --> 00:23:01.170 actively looking at this experimental AI guidance 00:23:01.170 --> 00:23:03.569 right now when they're analyzing active storms. 00:23:03.970 --> 00:23:06.430 That appears to be the case, based on the source 00:23:06.430 --> 00:23:09.660 material. It's being treated as another tool, 00:23:10.059 --> 00:23:12.099 another piece of guidance for them to consider 00:23:12.099 --> 00:23:15.180 in their complex forecast process. What's the 00:23:15.180 --> 00:23:17.359 hope there? The hope, as the source puts it, 00:23:17.480 --> 00:23:19.960 is that having access to this potentially very 00:23:19.960 --> 00:23:23.099 skillful AI data will provide valuable insights. 00:23:23.460 --> 00:23:25.799 Insights that could help the NHC forecasters 00:23:25.799 --> 00:23:28.119 refine their official forecasts, maybe improve 00:23:28.119 --> 00:23:30.940 the accuracy, or potentially provide earlier 00:23:30.940 --> 00:23:33.099 or more confident warnings about the hazards. 00:23:33.500 --> 00:23:36.450 It's a really critical phase of real -world, 00:23:36.490 --> 00:23:38.990 real -time evaluation. So the performance of 00:23:38.990 --> 00:23:41.650 this experimental model this season, under the 00:23:41.650 --> 00:23:44.809 eyes of the NHC experts, is really make or break 00:23:44.809 --> 00:23:46.950 for demonstrating its practical value. Absolutely. 00:23:46.950 --> 00:23:49.410 It's the ultimate test. Another major collaborator 00:23:49.410 --> 00:23:51.829 mentioned is CRI, the Cooperative Institute for 00:23:51.829 --> 00:23:53.829 Research in the Atmosphere. They're based at 00:23:53.829 --> 00:23:55.829 Colorado State University. Right. CSU does a 00:23:55.829 --> 00:23:57.910 lot of hurricane research. They do. And a team 00:23:57.910 --> 00:24:00.950 there, led by Dr. Kate Musgrave, conducted an 00:24:00.950 --> 00:24:04.089 independent evaluation of this AI model. An independent 00:24:04.089 --> 00:24:06.339 look. What did they find? According to the source, 00:24:06.660 --> 00:24:08.539 Dr. Musgrave's team concluded that the model 00:24:08.539 --> 00:24:11.339 demonstrated comparable or greater skill than 00:24:11.339 --> 00:24:14.119 the best operational models for track and intensity. 00:24:14.640 --> 00:24:17.359 So another group, looking independently, came 00:24:17.359 --> 00:24:19.880 to a similar conclusion as the internal Google 00:24:19.880 --> 00:24:23.019 tests. It seems so, yes. That kind of independent 00:24:23.019 --> 00:24:25.019 verification from a well -respected research 00:24:25.019 --> 00:24:27.980 institution like CIRA adds significant credibility 00:24:27.980 --> 00:24:29.960 to the initial findings. That's important. Do 00:24:29.960 --> 00:24:32.319 they say anything else? Yes. And it echoes the 00:24:32.319 --> 00:24:34.720 point about real time performance. Dr. Musgrave 00:24:34.720 --> 00:24:37.059 is quoted saying her team is looking forward 00:24:37.059 --> 00:24:39.599 to confirming those results from real time forecasts 00:24:39.599 --> 00:24:44.099 during the 2025 hurricane season. So again, reinforcing 00:24:44.099 --> 00:24:46.480 that while the historical testing is compelling, 00:24:46.940 --> 00:24:49.500 proving its reliability now in ongoing operational 00:24:49.500 --> 00:24:52.480 settings is paramount. Makes total sense. Compelling 00:24:52.480 --> 00:24:54.980 results in testing now under intense scrutiny 00:24:54.980 --> 00:24:57.500 in real time by the experts who live and breathe 00:24:57.500 --> 00:24:59.599 this stuff. Were there other partnerships mentioned? 00:24:59.940 --> 00:25:02.000 Yeah, the source also lists collaborations with 00:25:02.000 --> 00:25:04.259 the UK Met Office, the University of Tokyo in 00:25:04.259 --> 00:25:06.519 Japan, a private Japanese company called Weather 00:25:06.519 --> 00:25:09.759 News Inc, and other sort of trusted testers globally. 00:25:10.079 --> 00:25:11.960 So getting feedback from different regions too. 00:25:12.059 --> 00:25:15.099 Exactly. And that's important because tropical 00:25:15.099 --> 00:25:17.460 cyclones can behave differently in different 00:25:17.460 --> 00:25:20.980 ocean basins due to varying environmental conditions. 00:25:21.799 --> 00:25:24.180 Testing and getting feedback from experts dealing 00:25:24.180 --> 00:25:27.019 with typhoons in the Pacific or cyclones in the 00:25:27.019 --> 00:25:29.440 Indian Ocean provides a more complete picture 00:25:29.440 --> 00:25:31.980 of the model's strengths and weaknesses globally. 00:25:32.269 --> 00:25:34.710 So the overall goal of all this collaboration 00:25:34.710 --> 00:25:37.509 is? It's about gathering crucial feedback, as 00:25:37.509 --> 00:25:40.349 the source emphasizes. Feedback from the weather 00:25:40.349 --> 00:25:42.829 agencies, the emergency managers, the people 00:25:42.829 --> 00:25:44.930 who have to make life or death decisions based 00:25:44.930 --> 00:25:47.809 on forecasts. They need to understand how this 00:25:47.809 --> 00:25:50.670 new AI technology can best fit into the existing 00:25:50.670 --> 00:25:53.630 forecast process, how it can genuinely enhance 00:25:53.630 --> 00:25:56.009 the official forecasts, and ultimately how it 00:25:56.009 --> 00:25:58.230 can help inform better decisions to save lives 00:25:58.230 --> 00:26:00.829 and protect property. It's about bridging the 00:26:00.829 --> 00:26:03.549 gap from research to real world operational impact. 00:26:03.890 --> 00:26:05.809 It definitely sounds like the focus is on practical 00:26:05.809 --> 00:26:08.890 application, getting this tool refined and potentially 00:26:08.890 --> 00:26:11.589 integrated where it can do the most good. How 00:26:11.589 --> 00:26:13.450 are they making this technology more broadly 00:26:13.450 --> 00:26:15.890 accessible for experts to look at and evaluate? 00:26:16.119 --> 00:26:18.200 That's where a platform called Weather Lab comes 00:26:18.200 --> 00:26:21.339 in. The source describes it as an interactive 00:26:21.339 --> 00:26:24.160 website that Google DeepMind and Google Research 00:26:24.160 --> 00:26:26.559 have launched. It seems to be the main public 00:26:26.559 --> 00:26:29.220 -facing portal, specifically for the expert community, 00:26:29.720 --> 00:26:32.099 where they share these experimental AI weather 00:26:32.099 --> 00:26:35.269 models. Okay, Weather Lab. What does it actually 00:26:35.269 --> 00:26:37.710 offer? What can experts find there? It provides 00:26:37.710 --> 00:26:40.509 access to both live real -time predictions and 00:26:40.509 --> 00:26:43.089 also historical forecast archives from several 00:26:43.089 --> 00:26:45.990 of Google's experimental AI weather models. This 00:26:45.990 --> 00:26:48.369 includes the specific tropical cyclone model 00:26:48.369 --> 00:26:50.930 we've been focusing on, but also others like 00:26:50.930 --> 00:26:53.009 Weather Next Graph and Weather Next Gen, which 00:26:53.009 --> 00:26:55.269 might have different strengths or focuses. So 00:26:55.269 --> 00:26:58.269 they can see the AI output. Yes. And crucially, 00:26:58.549 --> 00:27:00.750 Weather Lab also displays predictions from well 00:27:00.750 --> 00:27:03.109 -established physics -based models right alongside 00:27:03.109 --> 00:27:05.690 the AI ones. They specifically mentioned showing 00:27:05.690 --> 00:27:09.250 output from ECMWF's ENS ensemble, the European 00:27:09.250 --> 00:27:11.549 model we talked about, for direct comparison. 00:27:11.809 --> 00:27:14.230 Ah, so you can see the AI forecast next to the 00:27:14.230 --> 00:27:16.289 traditional model forecast for the same storm. 00:27:16.509 --> 00:27:20.170 Exactly. That side -by -side comparison is critical. 00:27:20.609 --> 00:27:23.549 It allows experts from weather agencies, research 00:27:23.549 --> 00:27:26.829 groups, private companies, wherever, to directly 00:27:26.829 --> 00:27:29.589 evaluate the AI model's performance against the 00:27:29.589 --> 00:27:32.390 models they already know, use, and trust. They 00:27:32.390 --> 00:27:34.650 can see how it behaves, where it agrees, where 00:27:34.650 --> 00:27:36.769 it differs, and start to build an understanding 00:27:36.769 --> 00:27:39.210 of its characteristics. So it's a platform for 00:27:39.210 --> 00:27:41.490 transparency and comparative evaluation within 00:27:41.490 --> 00:27:43.529 the meteorological community. That seems to be 00:27:43.529 --> 00:27:46.299 a primary goal, yes. And transparency is further 00:27:46.299 --> 00:27:48.920 supported because, as the source highlights, 00:27:49.440 --> 00:27:52.079 Weather Lab also provides an archive of historical 00:27:52.079 --> 00:27:54.160 predictions from these models, apparently going 00:27:54.160 --> 00:27:56.579 back over two years. An archive, so you can look 00:27:56.579 --> 00:27:59.079 back at past storms. Yeah, and researchers or 00:27:59.079 --> 00:28:01.299 other experts can actually download this historical 00:28:01.299 --> 00:28:03.660 forecast data. This allows them to conduct their 00:28:03.660 --> 00:28:06.160 own independent, in -depth evaluations of a model's 00:28:06.160 --> 00:28:08.339 performance over longer periods, across different 00:28:08.339 --> 00:28:10.859 storm situations, and in all the various ocean 00:28:10.859 --> 00:28:13.779 basins around the world. That open access to 00:28:13.779 --> 00:28:15.380 historical data is really important. important 00:28:15.380 --> 00:28:18.220 for building wider scientific confidence and 00:28:18.220 --> 00:28:20.599 validating the technology thoroughly. OK, so 00:28:20.599 --> 00:28:23.420 it's a place to see live runs, compare models, 00:28:23.900 --> 00:28:26.359 and download historical data for deeper analysis. 00:28:27.119 --> 00:28:30.480 How does using this platform actually help, say, 00:28:30.680 --> 00:28:33.259 an emergency manager or a forecaster on a practical 00:28:33.259 --> 00:28:35.849 level when a storm is brewing? Well, the source 00:28:35.849 --> 00:28:37.970 suggests that by using Weatherlad to explore 00:28:37.970 --> 00:28:39.910 and compare the predictions from the different 00:28:39.910 --> 00:28:42.450 models, both AI and physics -based forecasters 00:28:42.450 --> 00:28:44.809 can get a more holistic view of the situation. 00:28:44.869 --> 00:28:47.029 Looking at everything together. Right. Seeing 00:28:47.029 --> 00:28:50.289 the AI model's output, maybe those 50 probabilistic 00:28:50.289 --> 00:28:53.289 scenarios it generates, alongside the ensemble 00:28:53.289 --> 00:28:56.170 plumes from traditional models like ENS, and 00:28:56.170 --> 00:28:58.230 comparing that with the latest satellite imagery 00:28:58.230 --> 00:29:01.049 and aircraft data, it gives a more robust understanding 00:29:01.049 --> 00:29:03.230 of the potential track, the potential intensity 00:29:03.230 --> 00:29:05.579 range, and importantly, uncertainty surrounding 00:29:05.579 --> 00:29:08.059 the forecast. A richer picture to inform decisions. 00:29:08.319 --> 00:29:10.680 Exactly. The source notes this helps agencies 00:29:10.680 --> 00:29:13.440 with better preparation. They can develop contingency 00:29:13.440 --> 00:29:16.019 plans based on the range of plausible scenarios, 00:29:16.079 --> 00:29:18.799 not just a single forecast line. It can help 00:29:18.799 --> 00:29:20.880 improve how they communicate the risks to the 00:29:20.880 --> 00:29:23.559 public talking about likelihoods and possibilities 00:29:23.559 --> 00:29:26.259 rather than just one outcome. And ultimately, 00:29:26.400 --> 00:29:28.680 it supports more informed decision -making about 00:29:28.680 --> 00:29:31.299 everything from evacuation orders to where to 00:29:31.299 --> 00:29:33.980 pre -position emergency crews and supplies. It's 00:29:33.980 --> 00:29:36.279 a decision support tool. It sounds like a potentially 00:29:36.279 --> 00:29:39.619 very powerful in addition to the forecasters 00:29:39.619 --> 00:29:41.799 toolkit for improving situational awareness. 00:29:42.740 --> 00:29:44.819 But, and this feels like a really important but, 00:29:45.079 --> 00:29:47.380 these are still experimental models, right? What 00:29:47.380 --> 00:29:49.720 are the caveats here? Absolutely critical point, 00:29:49.819 --> 00:29:51.759 and the source material is very clear about this, 00:29:51.759 --> 00:29:53.720 thankfully. It includes a prominent and essential 00:29:53.720 --> 00:29:56.599 disclaimer. Weather lab is a research tool. Okay, 00:29:56.660 --> 00:29:59.240 say that again, a research tool. Yes, not an 00:29:59.240 --> 00:30:01.539 official source for warnings. The live predictions 00:30:01.539 --> 00:30:03.500 shown on weather lab are generated by models 00:30:03.500 --> 00:30:05.619 that are still explicitly described as being 00:30:05.619 --> 00:30:08.900 under development. Experimental. Not finalized. 00:30:09.019 --> 00:30:11.240 Yes. Not operationally certified in the formal 00:30:11.240 --> 00:30:14.660 sense. Precisely. These are experimental models. 00:30:15.480 --> 00:30:17.980 Their output should never be taken or used by 00:30:17.980 --> 00:30:21.480 the public, or anyone really, as official weather 00:30:21.480 --> 00:30:24.500 forecasts or warnings. This cannot be emphasized 00:30:24.500 --> 00:30:27.559 enough, especially for people in hurricane -prone 00:30:27.559 --> 00:30:30.480 areas who might stumble across this site. It 00:30:30.480 --> 00:30:32.660 really can't. The source makes it absolutely 00:30:32.660 --> 00:30:35.920 explicit. For official weather forecasts and 00:30:35.920 --> 00:30:38.279 warnings that you should use to make safety decisions, 00:30:38.700 --> 00:30:41.940 you must refer to your designated local meteorological 00:30:41.940 --> 00:30:44.599 agency or your national weather service. Like 00:30:44.599 --> 00:30:46.640 a National Hurricane Center in the U .S. Exactly, 00:30:46.720 --> 00:30:49.259 or the equivalent official body in whatever country 00:30:49.259 --> 00:30:52.099 you're in. While these AI models are exciting 00:30:52.099 --> 00:30:55.140 and clearly valuable tools for experts to evaluate 00:30:55.140 --> 00:30:57.180 and potentially incorporate into their thinking, 00:30:57.319 --> 00:30:59.960 they are not the final authoritative word for 00:30:59.960 --> 00:31:02.740 public safety. The official forecasts and warnings 00:31:02.700 --> 00:31:05.180 still come from the human experts at the recognized 00:31:05.180 --> 00:31:09.319 agencies. Yes. Those experts synthesize all the 00:31:09.319 --> 00:31:11.900 available information data from multiple traditional 00:31:11.900 --> 00:31:15.359 models, ensemble systems, satellite observations, 00:31:16.000 --> 00:31:17.799 aircraft reconnaissance, their own experience 00:31:17.799 --> 00:31:20.779 and expertise, and now, potentially, insights 00:31:20.779 --> 00:31:24.119 from these experimental AI tools to issue the 00:31:24.119 --> 00:31:27.299 definitive public advisories and warnings. Their 00:31:27.299 --> 00:31:30.299 responsibility rests with them. That's a vital 00:31:30.299 --> 00:31:32.980 distinction. The potential is huge, the research 00:31:32.980 --> 00:31:36.539 is exciting, but always, always rely on the official 00:31:36.539 --> 00:31:38.660 sources for safety decisions. Couldn't say it 00:31:38.660 --> 00:31:41.640 better. The AI is adding new data points, potentially 00:31:41.640 --> 00:31:44.059 very valuable ones, to help inform those human 00:31:44.059 --> 00:31:47.000 experts, but it doesn't replace them or the official 00:31:47.000 --> 00:31:49.420 warning system. Okay. So let's try and pull all 00:31:49.420 --> 00:31:51.599 this together then. Thinking about the big picture, 00:31:51.819 --> 00:31:53.660 what's the real significance of this development? 00:31:53.759 --> 00:31:56.660 Why should you, our listener, care about this 00:31:56.660 --> 00:32:00.000 experimental AI hurricane model? Well, if we 00:32:00.000 --> 00:32:03.019 recap the main points, this AI model shows real, 00:32:03.299 --> 00:32:05.599 tangible promise in tackling that long -standing 00:32:05.599 --> 00:32:08.099 challenge in hurricane forecasting, getting both 00:32:08.099 --> 00:32:10.980 the track and the intensity right with high accuracy 00:32:10.980 --> 00:32:13.180 from a single system. Overcoming that tradeoff. 00:32:13.380 --> 00:32:15.660 Right. The internal tests they presented showed 00:32:15.660 --> 00:32:18.680 results that were, frankly, remarkable, especially 00:32:18.680 --> 00:32:21.579 that potential 1 .5 -day leap forward in track 00:32:21.579 --> 00:32:24.039 forecast lead time accuracy at the five -day 00:32:24.039 --> 00:32:26.700 mark. That's the kind of jump that historically 00:32:26.700 --> 00:32:29.559 has taken many years, even a decade or more, 00:32:29.660 --> 00:32:32.119 of incremental progress with traditional methods. 00:32:32.319 --> 00:32:34.500 And it wasn't just track. It also showed better 00:32:34.500 --> 00:32:36.720 intensity performance than a specialist model 00:32:36.720 --> 00:32:39.740 in those tests. Correct. And maybe, most importantly, 00:32:39.880 --> 00:32:42.220 this isn't just happening in a lab anymore. Leading 00:32:42.220 --> 00:32:45.319 operational forecast like the NHC and top research 00:32:45.319 --> 00:32:47.980 partners like CRA, they are actively looking 00:32:47.980 --> 00:32:50.440 at this model's performance right now in real 00:32:50.440 --> 00:32:53.480 time during this current hurricane season. The 00:32:53.480 --> 00:32:56.140 validation process is underway in the real world. 00:32:56.599 --> 00:32:58.619 So connecting that back to the listener, back 00:32:58.619 --> 00:33:01.140 to that $1 .4 trillion damage figure, back to 00:33:01.140 --> 00:33:03.660 the human impact, why does this specific development 00:33:03.660 --> 00:33:06.190 matter? It matters because of the direct link 00:33:06.190 --> 00:33:08.430 between forecast improvements and real -world 00:33:08.430 --> 00:33:11.710 outcomes. Even seemingly small gains in accuracy, 00:33:11.970 --> 00:33:14.150 getting the track just a bit better, predicting 00:33:14.150 --> 00:33:16.950 rapid intensification just a little sooner, or 00:33:16.950 --> 00:33:19.569 extending the reliable forecast lead time by 00:33:19.569 --> 00:33:22.849 half a day or a full day. Those improvements 00:33:22.849 --> 00:33:25.630 can have massive positive consequences. How so? 00:33:25.769 --> 00:33:28.589 better preparedness, more time for effective 00:33:28.589 --> 00:33:31.589 evacuations, which saves lives, more efficient 00:33:31.589 --> 00:33:34.170 staging of emergency resources, getting crews, 00:33:34.269 --> 00:33:36.589 supplies, shelters ready in the right places, 00:33:37.009 --> 00:33:39.549 better communication of risk to the public. It 00:33:39.549 --> 00:33:42.130 all contributes to reducing the devastating loss 00:33:42.130 --> 00:33:44.589 of life and the enormous economic disruption 00:33:44.589 --> 00:33:47.569 these storms cause. This AI model, based on its 00:33:47.569 --> 00:33:49.849 initial promise, holds the potential to deliver 00:33:49.849 --> 00:33:52.130 those kinds of tangible improvements. In this 00:33:52.130 --> 00:33:53.869 whole development, it feels like it fits into 00:33:53.869 --> 00:33:55.670 a bigger picture, a larger friend we're seeing 00:33:55.670 --> 00:33:58.450 with AI, doesn't it? Oh, absolutely. This is 00:33:58.450 --> 00:34:00.910 a really compelling example of how artificial 00:34:00.910 --> 00:34:03.190 intelligence is increasingly being applied to 00:34:03.190 --> 00:34:06.730 tackle really complex scientific problems and 00:34:06.730 --> 00:34:09.750 major societal challenges. The source material 00:34:09.750 --> 00:34:12.510 itself frames this work under Google's broader 00:34:12.510 --> 00:34:15.309 umbrella of AI for climate and sustainability. 00:34:15.550 --> 00:34:17.989 We're seeing AI used in drug discovery, finding 00:34:17.989 --> 00:34:20.929 new materials. Exactly. Optimizing energy grids, 00:34:21.289 --> 00:34:23.829 understanding complex biological systems, and 00:34:23.829 --> 00:34:26.289 now significantly improving our ability to predict 00:34:26.190 --> 00:34:29.690 and prepare for extreme weather events, which, 00:34:29.849 --> 00:34:31.869 let's face it, is becoming ever more critical 00:34:31.869 --> 00:34:34.630 in the context of a changing climate. It is genuinely 00:34:34.630 --> 00:34:37.730 exciting, I think, to see this incredibly powerful 00:34:37.730 --> 00:34:40.250 technology being pointed at problems that have 00:34:40.250 --> 00:34:43.050 such a direct, tangible impact on people's lives, 00:34:43.429 --> 00:34:45.409 on community safety and resilience. It really 00:34:45.409 --> 00:34:48.030 is. It represents a powerful new tool, a new 00:34:48.030 --> 00:34:50.369 way of thinking being added to the scientific 00:34:50.369 --> 00:34:52.989 arsenal. It offers the potential to accelerate 00:34:52.989 --> 00:34:55.429 discovery and boost our predictive capability. 00:34:55.400 --> 00:34:57.619 for phenomena that have historically been very, 00:34:57.619 --> 00:35:00.159 very difficult to forecast accurately. Okay, 00:35:00.159 --> 00:35:03.059 so as we wrap up this deep dive into AI hurricane 00:35:03.059 --> 00:35:05.420 forecasting on meteorology matters, here's maybe 00:35:05.420 --> 00:35:07.519 a final thought for you, the listener, to mull 00:35:07.519 --> 00:35:10.780 over. Oh, for it. If, and it's still an if based 00:35:10.780 --> 00:35:13.940 on this season's performance, but if these experimental 00:35:13.940 --> 00:35:16.800 AI models continue to prove their worth in real 00:35:16.800 --> 00:35:19.420 -time evaluations, how might this fundamentally 00:35:19.420 --> 00:35:22.320 change the dynamic inside? the forecast offices, 00:35:22.539 --> 00:35:24.920 the control rooms where those critical life -saving 00:35:24.920 --> 00:35:27.860 predictions are made. How do you see human expertise 00:35:27.860 --> 00:35:30.300 and AI guidance best working together in the 00:35:30.300 --> 00:35:33.800 future? Will the AI provide the baseline and 00:35:33.800 --> 00:35:37.260 the human fine -tunes or something else? That's 00:35:37.260 --> 00:35:39.119 the huge question, isn't it? The human -machine 00:35:39.119 --> 00:35:41.360 collaboration aspect. And maybe take it a step 00:35:41.360 --> 00:35:44.300 further. Could this... approach this idea of 00:35:44.300 --> 00:35:46.460 learning extremely complex dynamics directly 00:35:46.460 --> 00:35:49.280 from vast data sets, could that be applied effectively 00:35:49.280 --> 00:35:52.139 to predicting other chaotic, dangerous natural 00:35:52.139 --> 00:35:54.460 events that we still struggle with? Things like 00:35:54.460 --> 00:35:56.980 a major earthquake aftershock sequences or the 00:35:56.980 --> 00:35:59.480 unpredictable behavior of large wildfires. Those 00:35:59.480 --> 00:36:01.699 are fascinating and really important questions 00:36:01.699 --> 00:36:04.820 about where this technology might lead. Predicting 00:36:04.820 --> 00:36:08.300 complex natural systems is one of science's grand 00:36:08.300 --> 00:36:11.079 challenges. It seems like the potential is there. 00:36:11.320 --> 00:36:14.340 But figuring out the right way to integrate it, 00:36:14.559 --> 00:36:17.199 validate it, and rely on it is the next big step. 00:36:17.400 --> 00:36:19.300 Absolutely. The developments we've discussed 00:36:19.300 --> 00:36:21.719 today are definitely a significant stride forward. 00:36:22.039 --> 00:36:24.099 They open up really exciting new possibilities 00:36:24.099 --> 00:36:26.800 for prediction, but they also highlight just 00:36:26.800 --> 00:36:29.880 how important that careful ongoing validation 00:36:29.880 --> 00:36:32.619 processes, how crucial it is to figure out the 00:36:32.619 --> 00:36:35.760 best ways to integrate these tools, and importantly, 00:36:36.099 --> 00:36:38.559 they underscore the continuing irreplaceable 00:36:38.559 --> 00:36:41.199 role of human expertise in interpreting the outputs, 00:36:41.679 --> 00:36:43.960 understanding the limitations, and making the 00:36:43.960 --> 00:36:47.380 final responsible forecast decisions. Couldn't 00:36:47.380 --> 00:36:49.320 agree more. Well, that has certainly been an 00:36:49.320 --> 00:36:51.059 enlightening look at a potentially transformative 00:36:51.059 --> 00:36:53.320 technology in the weather world. Thank you for 00:36:53.320 --> 00:36:55.380 joining us for this deep dive on Meteorology 00:36:55.380 --> 00:36:57.519 Matters. My pleasure. It's a fascinating field 00:36:57.519 --> 00:37:00.039 right now. Hey, if you want to keep up with the 00:37:00.039 --> 00:37:02.800 latest in weather forecasting, especially tropical 00:37:02.800 --> 00:37:05.699 cyclones, you should definitely follow meteorologist 00:37:05.699 --> 00:37:08.500 Rob Jones. He shares great insights. Where can 00:37:08.500 --> 00:37:10.719 people find him? You can find him on Instagram. 00:37:11.000 --> 00:37:14.000 His handle is at Meteorologist. On TikTok, he's 00:37:14.000 --> 00:37:16.900 at TV Meteorologist. And over on YouTube, just 00:37:16.900 --> 00:37:20.320 search for Rob Jones Hurricane, R -O -B -J -O 00:37:20.320 --> 00:37:22.719 -N -E -S Hurricane. You'll find his channel there. 00:37:22.960 --> 00:37:25.099 And he also has a playlist for all the Meteorology 00:37:25.099 --> 00:37:27.260 Matters podcast discussions. Good stuff. Worth 00:37:27.260 --> 00:37:30.139 checking out. Definitely. All right. That's all 00:37:30.139 --> 00:37:32.019 for this time. Until our next deep dive, stay 00:37:32.019 --> 00:37:34.840 informed, stay curious, and please stay safe.