字幕列表 影片播放 列印英文字幕 On September 11th, 2023, weather predictions in the northeast of the U.S. sounded like this. All eyes are on Hurricane Lee. The storm has strengthened back to a category 3. We're expecting it to make a northward turn over the next few days. By September 16th, after being downgraded, storm Lee made landfall in Nova Scotia, Canada, flooding roads, downing trees and cutting out power for tens of thousands of people along the East Coast. At least five days before Hurricane Lee struck land, weather forecasts had roughly predicted its trajectory. But another forecast beat them three days before weather stations, an AI model created by Google predicted the cyclone's path. The AI revolution has reached meteorology, and it's at a time when we are responding to extreme weather more than ever. We're about to find out if it will help us prepare by bringing the future into clear view. Predicting future weather more than a few hours out starts with creating a snapshot of Earth's current atmosphere. Scientists do that by collecting data from sources like satellites and weather stations and buoys located around the world, taking pictures of clouds and measuring temperature and pressure and wind speed and humidity. All that disparate data gets fed into computers, which generate a 3D grid of boxes that represent the atmosphere, both vertically and horizontally. Computers then do a lot of physics to determine how these conditions interact with each other, and they produce a forecast. I think any forecast has like 150 trillion calculations. It's pretty amazing. All that math requires some of the world's most powerful supercomputers. The two big ones are run by the European Centre for Medium-Range Weather Forecasts and the National Weather Service in the U.S. To make a local weather forecast from this global model, meteorologists zoom in and refine their own forecasts with their local expertise. Like if they live in a hilly area or a flat area or near a lake, they'll adjust those models and do their own professional interpretation based on their area. No matter what, this initial 3D grid of that sphere is never going to exactly replicate reality. There's too many gaps in the data we can measure. That means forecasts get blurrier the further out you go. Which is why the big weather centers don't just generate one forecast. They tweak the initial data and produce up to 50 forecasts. It's called ensemble forecasting, and it helps meteorologists measure uncertainty. If all 50 forecast looks similar, there's a higher certainty in the prediction. But if there's a lot of variation there's much less. We got to kind of keep an eye to the sky. There's a potential of another storm in the works. This went from 0% chance to 40% at two o'clock, 70% at 8:00, and at 11:00. We have a tropical storm. Weather centers only release their forecasts every six hours. Because that's all today's computing power will allow. But what if that limit didn't exist? Before we explain that, we'll hear from the sponsor of this video. This episode is presented by Microsoft Copilot for Microsoft 365, your AI assistant at work. Copilot can help you solve your most complex problems at work, going far beyond simple questions and answers. From getting up to speed on a teams meeting in seconds to helping you start a first draft faster in Word, Copilot for Microsoft 365 gives everyone an AI assistant at work in their most essential apps to unleash creativity, unlock productivity, and uplevel skills. And it's all built on Microsoft's comprehensive approach to security, privacy, compliance, and responsible AI. Microsoft does not influence the editorial process of our videos, but they do help make videos like this possible. To learn more, you can go to microsoft.com/copilotforwork. Now back to our video. In 2020, a group of researchers published a data set called Era5 five, which contained about 40 years of the Earth's hourly weather data. The data set was just primed for using AI because it was huge. The data is nice and smooth, there's no missing values, and it's free just to step in. And say, "Hey, you have terabytes of data? I can learn how weather moves." AI models learned how weather moves, not through applying trillions of physics equations to the Earth's atmosphere, but by being trained on Era5's enormous historical data set. Researchers gave the models a snapshot of weather conditions, ask them to make a prediction, and then scored them on how closely that prediction matched what really happened. After a while, the models eventually got really good at this. By 2023, the tech companies Google, Huawei and Nvidia had developed models that rivaled traditional forecasting on variables like surface temperature, humidity and wind speed, and on some extreme weather events like the paths of tropical cyclones, atmospheric rivers, and extreme temperatures. These AI models still rely on the same observation data from the big weather centers, the data that creates that initial 3D grid snapshot. But they don't require anything close to six hours to produce a prediction. Huawei's PanguWeather model, for example, can produce a week long forecast in 1.4 seconds. Which means that we spent over a century figuring out the physics, the atmospheric science and the computational skill to bring us our modern day weather forecast. And now suddenly we have these AI models that have come out of, you know, the past 2 or 3 years, and they're getting the same skill, and now they run on a modest laptop. Despite the impressive results from these first AI models, there's still lots of work to do. Google's GraphCast predicted Hurricane Lee's path faster than traditional models, but it didn't prove it could predict a hurricane's intensity, which is a trickier calculation to make. These AI models are incentivized to get as many correct answers as they can through the scoring system. If you swing for it, swing for the fences, right? If it misses, the model is penalized very large. It says "No, you should never do that. Don't swing for the fences because the error is going to be huge." But by prioritizing safer, correct answers to boost the model's score, it could miss rare, outlier weather events. Plus, they are learning from 40 years of history, and historical weather has fewer extreme events than we do today or will have in the future due to climate change. But a big reason for optimism with these AI models comes from their ensemble forecasting. Instead of the traditional 50 ensemble forecasts, they can predict a thousand or more because they're freed from computing and time constraints. There's always going to be uncertainty in a weather prediction, but larger ensembles will help us measure that uncertainty better. That's an extremely useful context to say you are a emergency manager down in Florida who's dealing with the very difficult decision: Are you going to, you know, order an evacuation or not? You want as much information about the uncertainty as possible. Large ensembles might also catch a rare weather event that a 50-member ensemble would miss. Or, measure the probability of weather events even further into the future than our 10 day forecast. Are we magically going to get a crystal ball that lets us foresee perfectly into the future? Probably not. I think especially on like sub seasonal timescales, like multiple months out, we're going to be able to frame the statistical question with a lot more specificity and probably a much better quantification of the uncertainty. We do have a new winter storm warning. One thing we shouldn't expect to change any time soon is the role of the meteorologist... at least the ones you see on TV. If only because we fundamentally have to communicate uncertainty, and we have to walk through all the various what-ifs. And a human is the best tool that we have today to effectively communicate that and help somebody else make a decision. AI forecasting models are still in an experimental phase, but the European Centre for Medium-Range Weather Forecasts has started publishing AI forecast alongside their traditional ones for the public to compare. When we check the weather in the very near future, it might be powered by AI instead of physics based models or a combination of the two. And if we get things right, we'll have a sharper view of the weather events that we need to prepare for the most.