Artificial Intelligence could save lives by warning us about dangerous weather sooner than traditional forecasting systems. Early warning of extreme weather like storms, floods and heat waves, gives communities crucial time to prepare. Weather forecasts have become much more accurate over the years, and now AI's ability to rapidly analyse past events can take this to a new level. For example, GraphCast (created by Google) predicted where a recent hurricane would make landfall three days before existing methods. 

 

Traditional weather forecasting involves constantly taking measurements of hundreds of factors: air pressure, temperature, wind speed, humidity, etc.. These come from many sources including weather stations, satellites, balloons and buoys in the ocean. They even use readings that are taken by sensors on the noses of commercial airplanes. This mountain of data is fed into a supercomputer which can process quadrillions (a thousand trillion) of calculations every second. These complex equations are then used to simulate what will happen in the Earth's atmosphere in order to predict how the weather will change and evolve. This traditional method is very successful, but analyzing so much data takes a lot of time. So even with the help of supercomputers, the traditional forecast can be too slow to give a useful warning.

 

AI has a completely different approach that produces forecasts in less than one minute, while using a fraction of the computing power. It does not try to model every factor that is happening in the world. Instead, it has learned vast quantities of historical weather data, and it uses this knowledge to predict how current weather patterns will evolve. Because AI does not try to solve complex equations, it can make its forecasts very quickly. These are not as detailed as the traditional forecasts, but they are better at predicting severe events, and particularly at tracking the path of big storms. Experts say that AI models like GraphCast are intended to work alongside of traditional forecasts - not to replace them. “The rapid performance of AI models is crucial for predicting severe events quickly, but we still need the traditional approach to gather the information that trains the model.”