Today's weather forecasts are based on complex models that incorporate the laws that govern the dynamics of the atmosphere and oceans, and these models run on some of the most powerful supercomputers in existence. However, Alphabet (Google's parent company) has managed to predict global weather conditions for the next 10 days in just one minute using a single machine the size of a personal computer, thanks to artificial intelligence developed by DeepMind. The Google AI predicts the weather and this has only just begun.
In this article we are going to tell you how Google AI predicts the weather and how this technology has evolved.
Google AI predicts the weather
Surprisingly, this AI system outperforms most modern weather prediction systems in almost every aspect. Interestingly, it appears that this time artificial intelligence is serving as a complement to human intelligence rather than replacing it.
The European Center for Medium-Range Weather Forecasts (ECMWF) has an incredibly advanced system that underwent a major upgrade last year, improving its predictive capabilities. Hosted at its facilities in Bologna, Italy, There is a supercomputer equipped with approximately one million processors (in contrast to the two or four found in a personal computer) and an extraordinary computing power of 30 petaflops, equivalent to a staggering 30.000 trillion calculations per second.
This immense computational capacity is necessary for one of its tools, High Resolution Forecasting (HRES), which accurately predicts medium-term global weather patterns, which They generally span 10 days, with an impressive spatial resolution of nine kilometers. These predictions serve as the basis for weather forecasts delivered by meteorologists around the world. Recently, GraphCast, an artificial intelligence developed by Google DeepMind, has been used to measure the capabilities of this formidable system in weather prediction.
AI Study Results
The comparison results, published Tuesday in the journal Science, reveal that GraphCast outperforms HRES in predicting numerous weather factors. According to the study, Google's machine outperforms ECMWF's in 90,3% of the 1.380 metrics examined.
When focusing solely on the troposphere, the atmospheric layer where most weather events occur, and excluding data from the stratosphere, which is approximately 6 to 8 kilometers above the Earth's surface, artificial intelligence (A.I.) ) outperforms human-supervised supercomputers in 99,7% of cases. the variables analyzed. Surprisingly, this achievement was achieved using a machine that closely resembles a personal computer known as a tensor processing unit or TPU.
According to Álvaro Sánchez González, researcher at Google DeepMind, TPUs are specialized hardware that offers more efficient training and execution of artificial intelligence software compared to a normal PC, while maintaining a similar size. Just as a computer's graphics card focuses on rendering images, TPUs are designed to excel in matrix products. For GraphCast training, we used 32 TPUs over the course of several weeks. However, once the training is completed, a single TPU can generate predictions in less than a minute, as explained by Sánchez González, one of the creators of the device.
GraphCast and prediction systems
A notable distinction between GraphCast and existing prediction systems is its ability to incorporate historical data. The creators trained the system using meteorological data from the ECMWF archive dating back to 1979. This extensive data set covers the rainfall in Santiago and the cyclones that have impacted Acapulco over a period of 40 years. After a considerable amount of training, GraphCast has the remarkable ability to generate accurate weather predictions.
It only requires knowledge of weather conditions six hours before and immediately before your forecast to accurately predict the weather another six hours from now. Predictions are interdependent and each new forecast informs the previous one. Ferran Alet, co-creator of this impressive DeepMind machine, explains its inner workings: «Our neural network anticipates weather conditions six hours in advance. To forecast the weather in 24 hours, we simply evaluate the model four times. Alternatively, we could have trained separate models for the different time periods, such as one for six hours and one for 24 hours. However, "We understand that the underlying principles that govern weather remain consistent within a six-hour period."
"Therefore, if we can discover the appropriate 6-hour model and use its own predictions as input, we can accurately forecast the weather for the next 12 hours and repeat this process every six hours." According to Alet, this approach provides a substantial amount of data for a single model, resulting in more efficient training.
Until now, weather forecasts have been based on numerical weather prediction, which uses scientific equations developed throughout history to account for the various complexities of atmospheric dynamics. The researchers' findings establish a set of mathematical algorithms that supercomputers must run to generate predictions for the next few hours, days, or weeks (although reliability decreases significantly beyond 15 days). However, carrying out this task requires a very advanced supercomputer, which involves significant costs and extensive engineering efforts.
Google AI model predicts the weather
What is particularly notable is that these systems they do not use the weather conditions of the previous day or even the previous year, despite occurring in the same place and at the same time.
On the contrary, it approaches the task from a different angle, almost the opposite. Through its advanced deep learning capabilities, it uses extensive archives of past weather data to gain a comprehensive understanding of the intricate cause-and-effect dynamics that dictate the progression of Earth's climate.
According to José Luis Casado, spokesperson for the Spanish Meteorological Agency (AEMET), historical data is not taken into account in the atmospheric model. Casado clarifies that this model is based on existing observations and the most recent prediction made by the model itself. By accurately understanding the current state of the atmosphere, it is possible to forecast its future progression. Unlike machine learning techniques, this approach does not use historical data or predictions.
I hope that with this information you can learn more about Google's AI that predicts the weather and its characteristics.