The weather forecast may show AI storms ahead

The writer is author of ‘The Weather Machine’

On a misty evening several years ago, at the European Centre for Medium-Range Weather Forecasts, I watched one of the world’s most advanced computer weather models cycle through its twice-daily perambulation into the atmosphere of the near future.

While the screen in front of me flashed prosaic status messages, the supercomputer down the hall crunched trillions of variables each second. Over two hours, it used equations of physics to transform the most recent weather observations into a high resolution picture of the following 14 days of sky. This would become the raw material of meteorologists’ forecasts around the world — useful for choosing a jacket, flying a jetliner or seeking shelter from a storm. The modern forecast is an astonishing achievement of technology and scientific co-operation, painstakingly assembled over the past half-century.

So it came as something of a surprise this week when scientists at Google DeepMind published a paper in Science which suggests that the pre-eminence of this system might soon pass. Artificial intelligence may have brought us to the cusp of a new paradigm in weather forecasting. Rather than the laws of physics, Google’s GraphCast system is based on an analysis of 39 years of past weather. Rather than running on a supercomputer the size of a volleyball court, it runs on a single laptop. Rather than taking two hours, it takes less than a minute. And rather than “numerical weather prediction” — the technique behind the modern discipline — it is “machine learning-based weather prediction,” a fundamentally new approach with staggeringly precocious results. 

Do not expect your weather forecast to improve instantly. Do not expect Google to replace the Met Office. Well, not yet. But once again, AI is matching the efficacy of an existing technology with such breathtaking efficiency and alacrity that it has experts salivating at the possibilities. For the weather, that probably means more accuracy and more precision, farther out in time. This might even spill over from seasonal forecasts, measured in months, to the longer time spans of climate models.

In some ways, GraphCast’s methods are a return to the past. In 1922, the English mathematician Lewis Fry Richardson laid out a vision for weather forecasting. Richardson dreamt of calculating equations by the Norwegian meteorologist Vilhelm Bjerknes fast enough to be useful — ie faster than the future weather would arrive. Doing so would have required assembling a computing staff of 64,000 in a stadium, but it would have freed meteorologists from the ineffectiveness of matching past and present weather maps in order to guess at the future.

Richardson didn’t have Google but GraphCast was trained using a historical data set produced by ECMWF, essentially a continuous simulation of the atmosphere from 1979—2017. The model requires the two most recent states of Earth’s weather — the current time and six hours earlier — in order to predict the state of the weather six hours ahead. But while that work is staggeringly efficient, the data processing remains reliant on ECMWF’s substantial technical infrastructure.

It also requires the data to be collected in the first place, which is still done by fleets of satellites, hundreds of thousands of weather stations, and sundry other observation tools operated primarily by national meteorological organisations that have — for more than a century — exchanged that data as a matter of principle and policy. Now all that is potentially changing.

Indeed, the footnoted disclosures in the recent Science paper note two salient facts. This research was funded by Google DeepMind and Alphabet, and its Google-employed authors have filed a provisional patent application on its results. While that is expected and obvious, it marks a dramatic break with the history of weather prediction. The past few years have seen a boom in private observation — micro satellites, balloons and the adaptation of cell phone metadata to weather sensing. What it portends is a breakdown of the international order for weather that has held since the second world war. 

For the moment, GraphCast is still entirely dependent on this system’s output. But the fact that it is the creation of a private company breaks with tradition. Others have tried, but with less blinding success. The astonishing invention of Google’s search algorithm, among other technologies, has already broken numerous other industries. But weather is a different beast. A change to the current regime would be less about dismantling one industry in favour of another, and more about replacing a universal good with a private good.

There is no doubt value in better forecasts. What is different is that value will not, by stated purpose, be shared by all of us who live within the embrace of Earth’s atmosphere and, increasingly, under the threat of its ever more vociferous extremes.