This month's Splitblog topic comes from our development team lead, Bartosz. Anyone who knows him knows he is a human Swiss army knife and, as such, not only thoroughly versed in meteorology but also a gifted writer. So of course he insisted on writing the piece on his chosen topic himself. Read on!
Hey Siri, what's the weather going to be?
A simple question, really. But also one that seems not at all simple to answer. And looking closer, you start to notice everything that is needed to answer whether it is worth packing an umbrella or not.
First of all, we have to collect data. A great deal of it. Spread across the whole planet there are countless measuring stations recording temperature, air pressure and humidity, and wind direction and speed. On top of that we use data from weather balloons, aircraft, ships and satellites.
To determine the weather from that, we still need a weather model. This is not a physical model but a mathematical description of our weather through a multitude of equations and parameters into which we feed our measurements to see how the weather will develop.
Parameters? What parameters now? Well… that is a great deal of data and a great many equations, and until now calculating it required extremely powerful supercomputers – and even those could not manage the task fast enough, because what use is the best forecast if it is only ready once the weather has already happened?
So we set to work and simplify things, for example: what happens close to the ground? Or at the transition from water to land? The complicated processes are simplified considerably, making them a little less accurate but calculable in a reasonable time.
Enough of the weather service! This is supposed to be about AI. Or is there a connection?
Yes, there is. For some time there have been very exciting AI approaches that promise weather forecasts without being quite so computation- and time-intensive. Last year, Google's AI research division DeepMind published the model GraphCast, which, trained on historical data, needs only a fraction of the computing power of numerical weather models (that is, the kind described above) and can deliver a ten-day forecast in under a minute. And GraphCast appears to be not only faster but – at least in part – more precise than numerical weather models; it predicted earlier than numerical models where Hurricane Lee would make landfall.
It is hardly surprising, then, that all the major weather services are now experimenting with AI approaches, including the German Weather Service (DWD), which only a few days ago announced a breakthrough in AI-supported weather forecasting in a press release, with its newly developed model AI-Var.
At the pace the whole field of artificial intelligence is currently developing, it remains very exciting to see what the future holds – meteorologically too. And until then, perhaps we can be a little forgiving towards our meteorologists and their forecasts; as we have seen, the job is not exactly straightforward.
Sources: https://www.dwd.de/DE/presse/pressemitteilungen/DE/2024/20240617_pm_ki_news.html
https://www.spektrum.de/news/graphcast-neues-ki-modell-soll-genauere-wettervorhersage-liefern/2198859
