Machine learning models are taking over the field of weather forecasting, from a quick guess of “how long this rain will last” to a 10-day forecast to century-level forecasts. Technology is increasingly important to climate scientists and local apps and news channels – and yet it doesn’t “understand” the weather any more than you or I do.
For decades, meteorology and weather forecasting have been largely defined by integrating observations into carefully tuned physical models and equations. This is still true – there is no science without observation – but the vast archives of data have made it possible to create powerful AI models that cover just about any time scale you care about. And Google seeks to dominate the field today and for eternity.
At the short end of the spectrum, we have nowcasting, which is usually consulted for the question “do I need an umbrella?” This is served by DeepMind’s “nowcasting” modelswhich essentially looks at precipitation maps as a sequence of images – which they are – and attempts to predict how the shapes in those images will evolve and move.
With countless hours of Doppler radar to study, the model can get a pretty good idea of what’s going to happen next, even in fairly complex situations like a cold front bringing snow or freezing rain (like the showed Chinese researchers). build on Google’s work).
This model is an example of how accurate weather forecasts can be when made by a system that has no real knowledge of how that weather occurs. Meteorologists can tell you that when this weather phenomenon collides with another, you get fog, hail, or humid heat, because that’s what physics tells them. The AI model knows nothing about physics: being purely data-driven, it simply makes a statistical guess of what will come next. Just like ChatGPT doesn’t actually “know” what it’s talking about, weather models don’t “know” what they’re predicting.
This may come as a surprise to those who believe that a solid theoretical framework is necessary to produce accurate forecasts, and indeed scientists are still reluctant to blindly adopt a system that does not distinguish a drop of rain from a ray of sunlight. But the results are impressive nonetheless, and in low-stakes questions like “will it rain while I walk to the store,” it’s more than good enough.
Google researchers also recently presented a new, slightly longer-term model. called MetNet-3, which predicts up to 24 hours into the future. As you might guess, this brings in data from a larger area, like county or state weather stations, and forecasts on a larger scale. This concerns things like “will this storm cross the mountains or dissipate” and so on. Knowing whether wind speeds or heat are likely to enter dangerous territory tomorrow morning is essential for planning emergency services and deploying other resources.
Today brings a new development on a “medium” scale, in 7 to 10 days. Google DeepMind Researchers published an article in the journal Science describing GraphCastwhich “predicts weather conditions up to 10 days in advance more accurately and much faster than the industry’s benchmark weather simulation system.”
GraphCast zooms out not only in time but also in size, covering the entire planet at a resolution of 0.25 degrees longitude/latitude, or approximately 28 × 28 kilometers at the equator. This means predicting what the situation will be like at more than a million points around the Earth, and while some of these points are obviously of more obvious interest than others, the goal is to create a global system which accurately predicts the main weather conditions for the next week or so.
“Our approach should not be viewed as a replacement for traditional weather forecasting methods,” the authors write, but rather as “evidence that MLWP is capable of addressing the challenges of real-world forecasting problems and has the potential to complement and improve the best current methods. .”
It won’t tell you whether it will rain in your neighborhood or just the entire city, but it is very useful for larger-scale weather events like major storms and other dangerous anomalies. These occur in systems several thousand kilometers wide, which means that GraphCast simulates them in great detail and can predict their movements and qualities over days – and all using a single computing unit Google for less than a minute.
This is an important aspect: efficiency. “Numerical weather forecasts,” traditional physics-based models, are computationally expensive. Sure, they can predict faster than the weather, otherwise they’re worthless – but you have to have a supercomputer on the job, and even then it can take a while to make forecasts with slight variations.
Let’s say for example that you don’t know whether the intensity of an atmospheric river will increase or decrease before a cyclone crosses its path. You may want to make a few forecasts with different levels of increase, a few with different decreases, and one if it stays the same, so that when one of these eventualities occurs, you have the forecast ready. Again, this can be of critical importance when dealing with phenomena such as storms, floods and wildfires. Knowing a day earlier that you will need to evacuate an area can save lives.
These tasks can get very complex very quickly when you take into account many different variables, and sometimes you will need to run the model dozens or even hundreds of times to get a real sense of how things will turn out. If these predictions take an hour each on a supercomputer cluster, that’s a problem; if it’s a minute each on a desktop-sized computer, you have thousands of them, it’s not a problem at all – in fact, you might start thinking about predicting more variations and finer!
And that’s the idea behind the ClimSim project at AI2, the Allen Institute for Artificial Intelligence. What if you wanted to predict not just 10 different options for what next week might look like, but a thousand options for how the next century will unfold?
This type of climate science is important for all kinds of long-term planning, but with a huge amount of variables to manipulate and forecasts spanning decades, you can bet the computing power needed is just as enormous. So the AI2 team is working with scientists around the world to accelerate and improve these predictions using machine learning, thereby improving “forecasts” on a century scale.
ClimSim models work similarly to those discussed above: instead of plugging numbers into a manually tuned, physics-based model, they look at all data as an interconnected vector field. When one number increases and another increases by half, but a third decreases by a quarter, these relationships are embedded in the machine learning model’s memory even if it doesn’t know they relate. to (for example) atmospheric CO2, surface temperature and ocean biomass.
The project manager I spoke to said the models they built are impressively accurate while being much cheaper to make computationally. But he admitted that scientists, even if they keep an open mind, operate (as is natural) with skepticism. The code is all there if you want to take a look yourself.
With such long timescales and such a rapidly changing climate, it is difficult to find suitable ground truth for long-term forecasts, and yet these predictions become more valuable with each passing moment. And as GraphCast researchers have pointed out, this is not a replacement for other methods but a complementarity. There is no doubt that climate scientists will want every tool they can get.