The storm that has hit Galicia for a month is due to atmospheric water vapor flows, the forecast of which is crucial for meteorologists.
For years, artificial intelligence has been removing its creators, humans, from successive thrones. Now it was the turn of meteorology, one of the greatest human creations since the Roman augurs, and before that it opened the intestines of an animal to find out whether it was the ideal time to sow seeds or whether the next morning would be favorable for war. Current weather forecasts are made using very complex models based on the laws that govern the dynamics of the atmosphere and oceans, running on some of the most powerful supercomputers in the world. Now Alphabet (Google’s parent company) can predict the weather on the planet in 10 days in a minute using a single machine the size of a personal computer and the artificial intelligence of DeepMind. And it does so by outperforming the most modern weather forecasting systems in almost all respects. However, this time it appears that artificial intelligence is complementing rather than replacing human intelligence.
The European Center for Medium-Range Weather Forecasts (ECMWF) has such an advanced system. Last year he renewed his predictive power. At its location in Bologna, Italy, a supercomputer operates with around a million processors (compared to two or four in a personal computer) and a computing power of 30 petaflops, or around 30,000 billion calculations per second. And it takes so many flops for one of its tools, the High Resolution Forecast (HRES, in English), to be able to predict the weather on the entire planet in the medium term, usually 10 days, with great precision and to do so with a spatial one Resolution of nine kilometers. This is where men and women’s predictions about the weather across half the world come from. This Goliath was measured by GraphCast, Google DeepMind’s artificial intelligence for weather forecasting.
The results of the comparison, published today Tuesday in the journal Science, show that GraphCast predicts hundreds of meteorological variables equally or better than HRES. As they show, the Google machine outperforms the ECMWF machine in 90.3% of the 1,380 metrics considered. When the data related to the stratosphere, located about 6 to 8 kilometers high in the sky, is discarded and the analysis is limited to that of the troposphere, the atmospheric layer where the next meteorological events occur, artificial intelligence (AI) outperforms Supercomputing under the supervision of humans in 99.7% of the variables analyzed. This was achieved using a machine very similar to a personal computer called a Tensor Processing Unit (TPU).
After training, each prediction can be made in under a minute using a single TPU. [máquina] “Much more efficient than a normal PC, but of similar size”
Álvaro Sánchez González, DeepMind researcher and co-creator of GraphCast
“TPUs are specialized hardware for training and running artificial intelligence software, much more efficient than a regular PC but of similar size,” explains Google DeepMind researcher Álvaro Sánchez González. “Just as the computer’s graphics card (also called GPU) specializes in rendering images, TPUs specialize in producing matrix products. To train GraphCast, we use 32 of these TPUs over several weeks. However, any post-training prediction can be made in less than a minute with a single TPU,” explains Sánchez González, one of the device’s developers.
One of the big differences between GraphCast and current forecast systems is that it is based on weather history. Its creators trained it on all meteorological data stored in the ECMWF archive since 1979. These include both the rains that have fallen in Santiago since then and the hurricanes that have arrived in Acapulco in 40 years. It took a while to train it, but once it’s done, GraphCast only needs to know the weather six hours ago and the weather just before its new forecast is published, in order to know in a second what the weather will be in another six hours will be. And each new prediction is reflected in the previous one.
Ferran Alet, also from DeepMind and co-developer of the machine, explains in detail how it works: “Our neural network predicts the weather six hours in the future. If we want to predict the weather in 24 hours, we simply evaluate the model four times. Another option would have been to train different models, one for 6 hours, another for 24 hours. But we know that in 6 hours the physics will be the same as it is now. So we know that if we find the right 6-hour model and give it its own forecasts as input, it should predict the weather in 12 hours and we can repeat the process every six hours.” This gives them “a lot more data for a single model, making training more efficient,” concludes Alet.
Until now, predictions have been based on so-called numerical weather forecasting, which uses physical equations that science has provided throughout its history to respond to the various processes that make up a system as complex as the dynamics of the atmosphere. Their results define a series of mathematical algorithms that the supercomputers must execute in order to obtain the forecast for the next hours, days or weeks in a matter of minutes (although there are also longer-term algorithms that provide reliability). drops dramatically after 15 days). To do all this, the supercomputer has to be very good, which means enormous costs and a lot of engineering work. Perhaps the striking thing is that these systems do not take advantage of the weather that occurred at the same place and time yesterday or last year. GraphCast does it differently, almost in reverse. Its deep learning uses decades of historical weather data to learn a model of the cause-and-effect relationships that determine the evolution of Earth’s weather.
José Luis Casado, spokesman for the Spanish Meteorological Agency (AEMET), explains why historical data is omitted: “The atmospheric model uses the available observations and the immediate previous prediction of the model itself: if the current state of the atmosphere is well known, its future Development can be predicted. Unlike machine learning methods, it does not use predictions or historical data.
“The significance of DeepMind’s work is that it shows that the predictive prediction of traditional models can even be improved using artificial intelligence.”
Ignacio López Gómez, climate scientist at Google Research
From Google Research headquarters in California (USA), researcher Ignacio López Gómez develops weather forecasting systems based on extensive data. At the beginning of the year he published his latest work, in which he uses artificial intelligence to predict heat waves. Although he knows several of GraphCast’s creators, he was not involved in its design and calculations. “The importance of the work of DeepMind and other similar models (such as the Pangu Weather system recently developed by Chinese scientists) is that they show that the predictive forecasting of traditional models can be achieved or even improved by artificial intelligence,” admits López a Training the AI models is expensive, but once trained, they can make weather forecasts much more efficient. “Instead of requiring supercomputers, artificial intelligence-based predictions can even be made on personal computers in a reasonable time.”
The ECMWF has taken note of this and is already developing its own AI-based forecasting system. In October they announced that they already had the first alpha version of their AIFS (or Artificial Intelligence/Integrated Forecasting System). “It’s based on the same methodology as Google,” says AEMET’s Casado. “Although AIFS is not a fully functional system, it is a major step forward,” he adds. As the creators of GraphCast concluded in their academic article, AI is not a replacement for human ingenuity and certainly not for “traditional weather forecasting methods that have been developed over decades and rigorously tested in many real-world contexts.” In fact, ECMWF actively collaborated with Google by giving them access to the data and supporting them in this project. Casado concludes that “traditional models based on physical equations and new machine learning models based on data could be complementary.”
You can follow THEME on Facebook, X and Instagram, or sign up here to receive our weekly newsletter.