Previously, there were concerns about the AI's ability to accurately predict extreme weather, but the tests have bypassed the problem (Shutterstock).

For some time, weather forecasts were not respected, with many skeptics even saying that "the problem with weather forecasting is that it's often right to ignore it, and it's often wrong to rely on it."

This culture has been shaped by a long history of predictions that sometimes mismatched reality, prompting American comedian Rodney Dangerfield to once say that "weather forecasts are not respected." But there are high hopes that employing AI in this task will give more respect, as confirmed by a study by a research team from Google DeepMind.

In that study, published in the latest issue of the journal Science, the AI-powered program was announced as graphcast, several times faster than government models that have existed for decades, in which countries have invested hundreds of millions of dollars, as it can predict air pressure, temperature, humidity and wind for up to 10 days in advance.

Big tech companies, including Google, Microsoft, Nvidia and Huawei, have sought rapid progress over the past two years in weather forecasting using AI, all of which have published academic studies claiming that their AI models work at least like traditional weather models. In their latest study, however, Google researchers declared that it had exceeded the European model, which is widely considered the gold standard for prediction.

Bus sized computer

Traditional weather models, such as the European Model run by the European Center for Medium-Range Weather Forecasts in Reading, Britain, and the American one run by the National Oceanic and Atmospheric Administration, provide predictions based on complex mathematical equations, and these equations support such life-saving predictions and warnings around the world, but they are expensive to operate because they require huge amounts of computing power, accomplished by a supercomputer the size of a school bus.

AI models use a different approach to forecasting, first being trained to recognize weather patterns through vast amounts of historical weather data, and then generating predictions by understanding current conditions and applying what they have learned from historical patterns. This process is computationally less intensive and can be completed in minutes or even seconds on much smaller computers.

Applying to Google DeepMind's artificial intelligence model "Graphcast", the researchers announced in their study that it was trained on nearly 40 years of historical data, which helped it provide a 10-day forecast of 6-hour intervals for locations around the world in less than a minute, using a non-energy-consuming minicomputer, whereas in the traditional model it took an hour or more to accomplish the same feat. Using a supercomputer that consumes a large amount of energy.

AI limits the ability to interpret prediction reasons, preventing it from being used in advice to farmers (Base for Intellia)

Precision race. Other track

In addition to speed and cost, Graffcast also gained in accuracy, being 10% more accurate than the European model for more than 90% of the weather variables evaluated.

The performance of the new model was evaluated against the European model, not only for individual weather variables such as temperature, wind and pressure, but also in predicting extreme events including tropical cyclones, rivers (long, narrow corridors found in the atmosphere that can transport moisture for thousands of miles), heat waves and hail.

Although researchers have previously expressed concerns about the AI's ability to accurately predict extreme weather — in part because there are relatively few such events in the past that AI can learn from — Graffcast has bypassed this problem somewhat, reducing hurricane forecasting errors by 10 to 15 miles in a 2- to 4-day time interval, and improving predictions of river-associated water vapor by 10 percent. to 25%, more accurate forecasts of extreme heat and cold were made 5 to 10 days ahead of schedule.

Big tech companies have made rapid progress over the past two years in predicting the weather using artificial intelligence (Getty Images)

Deep Learning Networks. "Black Boxes"

These advantages of Graffcast don't mean it's better than current methods when it comes to highly local forecasts like the likelihood of rain in your area, but it does outperform forecasting weather events over larger areas such as tropical cyclones and unusual temperature fluctuations.

Remy Lam of Google DeepMind and lead author of the study says in a report published by Tech Explore: "Our approach should not be seen as a substitute for traditional methods of weather forecasting, which have been developed over decades and rigorously tested in many real-world contexts, and offer many features that we have not yet explored. Complement and improve the best current methods."

The director of the Central Laboratory for Agricultural Climate at the Egyptian Ministry of Agriculture, Mohamed Abd Rabbo, confirms that artificial intelligence will not replace at that stage the traditional methods of forecasting. He says in a telephone interview with "Al Jazeera Net": "Artificial intelligence models, despite their success, do not give the possibility to explain how they reach specific predictions, as some of these models - especially deep learning networks - are often considered as black boxes, while one of the main roles of forecasters is to interpret information."

"For example, in terms of weather implications for agriculture, I need to translate the information obtained from the forecasts into a recommendation that I communicate to the farmer well in advance of the climate phenomenon so that he can take preventive action, a concept known as climate-smart agriculture, and I will not be able to do this well without explanation, and predictions based on artificial intelligence models do not enable this to happen."

"One of the challenges is also that some AI models may inherit biases from training data or the algorithms used, leading to deviant predictions, so the quality of data in this direction is a very important factor for the success of AI models, and quality also includes the use of diverse and high-quality data sets, and this may not be achieved in some areas, where data may be limited or of poor quality, which affects the accuracy of predictions," Abed Rabbo explains.

Another challenge is that weather patterns are changing, and AI models need to adapt to evolving conditions, and ensuring they succeed is a major challenge.

Source : Al Jazeera + Agencies