Machine learning models are taking over the field of weather forecasting, from a quick “how long will this rain last” to a 10-day outlook, to level forecasts of the century. The technology is increasingly important to climate scientists as well as to apps and local news stations — but it doesn’t “understand” the weather any more than you or I do.
For decades, meteorology and weather forecasting have often been defined by fitting observations to carefully calibrated physics-based models and equations. That’s still true — there’s no science without observation — but massive data archives enable powerful AI models that cover any time scale you can care to take. And Google is looking to dominate the field from now on forever.
At the short end of the spectrum we have the immediate forecast, which is often consulted for the question “do I need an umbrella?” It is served by DeepMind’s “nowcasting” modelswhich basically looks at rain maps like a sequence of images – where they are – and tries to predict how the shapes of the images will develop and shift.
With countless hours of doppler radar to study, the model can get a solid idea of what will happen next, even in relatively complex situations such as a cold front that brings the snow or freezing rain (as shown by Chinese researchers. building the work of Google).
This model is an example of how accurate weather predictions can be when made by a system that has no actual knowledge of how the weather is happening. Meteorologists may tell you that if this climate event runs up against another, you’ll get fog, or hail, or humid heat, because that’s what physics tells them. The AI model knows nothing about physics – being purely data-based, it only makes statistical predictions about what will happen next. Just like ChatGPT doesn’t “know” what it says, weather models don’t “know” what they predict.
It may come as a surprise to those who think that a strong theoretical framework is needed to make accurate predictions, and indeed scientists are still wary of blindly adopting a system without knowing one. raindrops from sunshine. But the results are impressive nonetheless and on low stakes things like “is it going to rain while I’m walking to the store” it’s more than fine.
Google researchers also recently presented a new, somewhat longer term model called MetNet-3, which predicts up to 24 hours into the future. As you might guess, it brings in data from a larger area, such as weather stations across a county or state, and its predictions take place on a larger scale. This is for things like “will the storm cross the mountains or disappear” and so on. Knowing whether strong winds or heat are likely to enter danger territory tomorrow morning is important for planning emergency services and deploying other resources.
Now brings a new development in the “medium-range” scale, which is 7-10 days in the future. Google DeepMind researchers published an article in the journal Science describing GraphCastwhich “predicts weather conditions up to 10 days in advance more accurately and faster than the industry’s gold-standard weather simulation system.”
GraphCast zooms out not just in time but in size, covering the entire planet at a resolution of .25 degrees longitude/latitude, or about 28×28 kilometers at the equator. That means predicting what will happen at more than a million points around the Earth, and even if some of the points are of more obvious interest than others, the point is to create a global system that is accurate. which predicts the main weather patterns for. in the next week or so.
“Our approach should not be considered a replacement for traditional weather forecasting methods,” the authors wrote, but “is evidence that MLWP is able to address the challenges of real-world forecasting problems and has potential to complement and improve current best practices. .”
It won’t tell you if it’s going to rain in your neighborhood or just across town, but it’s very useful for larger weather events like major storms and other dangerous anomalies. This happens in systems thousands of kilometers across, meaning that GraphCast simulates them in sufficient detail and can predict their movements and quality for days – and all using a Google compute unit without one minute.
That is an important aspect: efficiency. “Numerical weather prediction,” traditional physics-based models, are computationally expensive. Of course they can predict faster than the weather, otherwise, it’s useless – but you need to get a supercomputer to work, and even then it can take a long time to make predictions with small differences.
Say for example you are not sure if an atmospheric river will increase or decrease in intensity before an incoming storm crosses its path. You can make some predictions with different levels of increase, and some with different decreases, and one if it stays the same, so that if one of the events happens, you will be ready with the forecast. Again, this can be very important when it comes to things like hurricanes, floods, and fires. Knowing a day in advance that you need to evacuate an area can save lives.
These jobs can get really complicated fast when you’re accounting for a lot of different variables, and sometimes you need to run the model many times, or hundreds, to see what happens. If those predictions come an hour each on a supercomputer cluster, that’s a problem; if it’s a minute each on a desktop-size computer with thousands, no problem — in fact, you can start thinking about predicting more and better differences!
And that’s the idea behind it the ClimSim project in AI2, the Allen Institute for Artificial Intelligence. What if you wanted to predict not just 10 different options for what next week will look like, but a thousand options for how the game will play out in the next century?
This kind of climate science is important for all kinds of long-term planning, but with so many variables to manipulate and forecasts coming out over decades, you can bet the computing power needed is just as great. That’s why the AI2 team is working with scientists around the world to accelerate and improve predictions using machine learning, developing “predictions” on a century scale.
ClimSim models work similarly to those mentioned above: instead of plugging numbers into a physics-based, hand-tuned model, they view all data as an interconnected vector field. If one number goes up and a reliable case of another goes up by half as much, but a third goes down by a quarter, those relationships are embedded in the machine learning model’s memory even if they don’t exist. it is known to be related to (say) atmospheric CO2, surface temperature, and ocean biomass.
The project lead I spoke with said the models they built were very accurate while being orders of magnitude cheaper to perform computationally. But he admits that scientists, while they keep an open mind, are operating (as is natural) from a place of skepticism. The code is all here if you want to see for yourself.
With such long periods of time, and with rapid climate change, it is difficult to find appropriate ground truth for long-term predictions, but those predictions are increasingly valuable to all time. And as the GraphCast researchers point out, it is not a substitute for other methods but a complement. No doubt climate scientists want every tool they can get.