Google DeepMind's Weather AI Can Forecast Extreme Weather Faster and More Accurately 40
In research published in Science today, Google DeepMind's model, GraphCast, was able to predict weather conditions up to 10 days in advance, more accurately and much faster than the current gold standard. From a report: GraphCast outperformed the model from the European Centre for Medium-Range Weather Forecasts (ECMWF) in more than 90% of over 1,300 test areas. And on predictions for Earth's troposphere -- the lowest part of the atmosphere, where most weather happens -- GraphCast outperformed the ECMWF's model on more than 99% of weather variables, such as rain and air temperature. Crucially, GraphCast can also offer meteorologists accurate warnings, much earlier than standard models, of conditions such as extreme temperatures and the paths of cyclones. In September, GraphCast accurately predicted that Hurricane Lee would make landfall in Nova Scotia nine days in advance, says Remi Lam, a staff research scientist at Google DeepMind. Traditional weather forecasting models pinpointed the hurricane to Nova Scotia only six days in advance.
[...] Traditionally, meteorologists use massive computer simulations to make weather predictions. They are very energy intensive and time consuming to run, because the simulations take into account many physics-based equations and different weather variables such as temperature, precipitation, pressure, wind, humidity, and cloudiness, one by one. GraphCast uses machine learning to do these calculations in under a minute. Instead of using the physics-based equations, it bases its predictions on four decades of historical weather data. GraphCast uses graph neural networks, which map Earth's surface into more than a million grid points. At each grid point, the model predicts the temperature, wind speed and direction, and mean sea-level pressure, as well as other conditions like humidity. The neural network is then able to find patterns and draw conclusions about what will happen next for each of these data points.
[...] Traditionally, meteorologists use massive computer simulations to make weather predictions. They are very energy intensive and time consuming to run, because the simulations take into account many physics-based equations and different weather variables such as temperature, precipitation, pressure, wind, humidity, and cloudiness, one by one. GraphCast uses machine learning to do these calculations in under a minute. Instead of using the physics-based equations, it bases its predictions on four decades of historical weather data. GraphCast uses graph neural networks, which map Earth's surface into more than a million grid points. At each grid point, the model predicts the temperature, wind speed and direction, and mean sea-level pressure, as well as other conditions like humidity. The neural network is then able to find patterns and draw conclusions about what will happen next for each of these data points.
Low tar, low nicotine (Score:2, Funny)
"Yes, we have less roach droppings than the other leading brands!" - George Carlin
That's what weather prediction has come down to. It's "more accurate"... but still wrong.
Re: (Score:2)
AI cannot beat Lyapunov exponents, no matter how much hype the tech companies manage to fart out.
Hurricane Spoiler Alert by AI (Score:2)
Interesting (Score:4, Insightful)
And, in terms of forecasting, very useful. But because nobody can pick out the physical equations governing the predictions, it can't tell us why those predictions are accurate or what the physical models neglected/incorrectly weighted.
Because we can't see the workings, we don't know how reliable it'll be in a warming climate where the patterns are shifting, or how to fix any unreliability that may arise from such issues.
It's like an intuitive mathematician - if you can't see the working, then you don't know if the intuition generalises or under what conditions.
Re: (Score:2)
Because the AI isn't using equations. It says that right there in the summary. It's using historical data and finding the patterns and applying them to current data. Using physical equations works if the equations never change. I'm willing to bet that with the way our climate is shifting, the equations to accurately predict are also shifting. However, there are probably many other patterns that we haven't noticed or detected that an AI can point out.
Re: (Score:2)
It was trained on decades of historical data in the form of a high resolution map of conditions over time over every point we had available.
Whether the AI's network has derived "physical equations" or not is basically unknowable.
The bigger problem with our bespoke physical simulations, is that they're only as good as the amount of data we have, and how well our model predicts what comes from that data. And ultimately, it's only as good as the people
Re: (Score:2)
Well, it's not really possible to say what it's using.
It was trained on decades of historical data in the form of a high resolution map of conditions over time over every point we had available.
Whether the AI's network has derived "physical equations" or not is basically unknowable.
Yes, it is. I guess you could say the patterns it finds are types of equations (and maybe they are, maybe they aren't), but either way, they could easily have the AI spit out essentially a dump log of everything it used to come up with its predictions.
Ultimately, I'd say the AI has derived better models than us, which is pretty much to be expected with enough training.
I'd argue that it isn't necessarily a better model since it's a completely different methodology. The AI isn't trying to run simulations using realistic physics equations. That's what the current meteorology software does and it's very computationally taxing.
Re: (Score:2)
Yes, it is. I guess you could say the patterns it finds are types of equations (and maybe they are, maybe they aren't), but either way, they could easily have the AI spit out essentially a dump log of everything it used to come up with its predictions.
No, that's not how neural networks work.
I'd argue that it isn't necessarily a better model since it's a completely different methodology. The AI isn't trying to run simulations using realistic physics equations. That's what the current meteorology software does and it's very computationally taxing. The simulations with real world physics can simulate situations which have never happened before, whereas the AI would have troubles with that since there is no data to work from. Now, which is more useful in the day to day predictions of weather? Well, probably the AI. However, the simulations can help prepare us for random rarer catastrophes that we don't otherwise have a significant amount of historical data for, and potentially even help feed data into the AI.
Again, you don't know the methodology.
A set of inputs was fed into a convoluted neural network with millions of trained calculations performed on that data. The model it represents isn't quite unknowable, but it is practically unknowable.
Re: (Score:2)
You seem to be missing something since what you are saying isn't relevant to what I am saying.
Re: (Score:2)
You and parent can be the first people alive to reverse engineer a 37 million parameter NN!
Stupid fucker.
Re: (Score:2)
A log of the 37.5 million parameters in the GNN used for this isn't going to be useful for shit.
You have no fucking idea what they're doing. You can look at the input, and you can look at the output, and try to model what it's doing, and in the case where you were using 37.5 million parameters to calculate 1+1, you'd probably come up with a good model of what it w
Re: (Score:2)
Lmao, except that WAS NOT what I was referring to, so you are still a complete fucking retard responding to what you made up in your own head.
Re: (Score:2)
Yes, it is. I guess you could say the patterns it finds are types of equations (and maybe they are, maybe they aren't), but either way, they could easily have the AI spit out essentially a dump log of everything it used to come up with its predictions.
This is so fucking ignorant I don't even know how to diplomatically call you stupid, because I'm not sure if it's just flat out ignorance, or if you're simply incapable of complex reasoning.
Pattern matching is certainly something that ANNs excel at, but ANNs do not just "match patterns".
They take input data, run it through a complex learned network of sigmoids that are trained to produce a result you want.
To "log" that would be laughably stupid. Hell, Google's paper even says that they can only
Re: (Score:1)
Lmao still talking shit and failing to realize. It must be difficult having zero critical thinking skills.
Re: (Score:2)
You can't provide a retort of any substance, because you know damn well you're in over your head and trying to deflect.
You're not fooling anyone.
Re: (Score:1)
Your dumb ass is still here?
Re: (Score:2)
lol, you kill me. What a sad fucking human you must be to speak with authority to shit you have so little fucking understanding of.
Re: (Score:2)
I could never imagine being such a condescending piece of shit who assumes he's the explaining the right things and then doubles and triples down on it.
Re: (Score:2)
All this time, if only we had known all you needed to do was have the NN dump the logs of what it used to come up with its answer... We've all been so stupid!
No fuckstick. You said some stupid fucking shit because you were ignorant, and now you're trying to claw your way out. Ain't happening.
Re: (Score:2)
Oooo, quadrupling down. Interesting move. Let's see if it pays off...
Nope.
Re: (Score:2)
Re: (Score:1)
If we could figure out why existing "predictions are accurate or what the physical models neglected/incorrectly weighted" the existing models would be better, wouldn't they?
Re: Interesting (Score:2, Insightful)
Re: (Score:2)
No, because it is a black box it gives no understanding of how the weather works and so does not qualify as a scientific theory, nor does it make any falsifiable predictions beyond those that fit the data set it was fed.
Re: (Score:1)
No, because it is a black box it gives no understanding of how the weather works and so does not qualify as a scientific theory, nor does it make any falsifiable predictions beyond those that fit the data set it was fed.
It's pretty amazing that software with "no understanding of how the weather works" (your words), beats out the science(TM) experts 90% of the time.
Which coincidently is exactly what the people who haven't fallen for the global warming hoax have said all along.
Something that doesn't know shit about the calculations beats the scientists(TM) "on more than 99% of weather variables".
You warmist clowns can go home now.
Re: (Score:2)
Global warming hasn't been a hoax for any of the 200 years scientists have studied it.
If you can't tell the difference between local weather and global climate, sell your geek card and 4-digit uid because you don't qualify.
If AI was all they say it is (Score:4, Interesting)
AI,
Would be used to game then stock market and make the AI inventor the richest entity on earth. The fact that Google is pissing away resources on predicting weather tells me AI is worthless.
Re: (Score:2)
Moreso because stock markets are volatile and subjective. Stock prices move based on subjective opinions of shareholders quite often. An AI isn't good at subjective opinions. An AI could be used to attempt to manipulate a market similar to those hedge funds morons who were purposely over shorting GameStop, however, as we saw it can be easily combated once noticed. One company puts an AI out to do this, another AI will pop up to counter it.
Plus, stock markets are an unnecessary waste of time to begin with.
Re: (Score:2)
The closer to random the less of a pattern can be found within the noise. What would be nice is a measurement of randomness that can be used and compared to the effective pattern matching of the AI which would give you a predictor of how well an AI could manage to find some at least short term pattern in the noise.
Stocks overall are pretty random.
Re: (Score:2)
Re:If AI was all they say it is (Score:4, Insightful)
AI,
Would be used to game then stock market and make the AI inventor the richest entity on earth. The fact that Google is pissing away resources on predicting weather tells me AI is worthless.
Lots of folks are already doing that.
The issue with using AI to predict a system like the stock market is to make money you need to find a price mismatch (stock should be worth $X but it's currently worth $Y). The problem is when you act on that information you push the price towards $X. Basically, if one investor has a great AI they make a killing, if a bunch have a great AI then the market already reflects their views and they just make small margins.
Basically it's the efficient market hypothesis [wikipedia.org].
Forecast Extreme Weather Faster and More Accuratel (Score:2)
I'd prefer if it could Forecast Normal Weather Extremely Accurately.
The Old Traditional Method Souped Up (Score:2)
Prior to the advent of computers powerful enough to run useful numerical weather prediction models this how weather forecasting was done. Huge archives of historical weather data were consulted to find weather patterns that matched the current ones, and provided the basis for the forecast. The immense pattern matching capabilities of current neural networks resurrects this approach.
It will be interesting to see if a hybrid approach can improve things further. As weather patterns change with global warming I
Biased assumptions (Score:2)
The big problem with relying only on historical data to make future predictions is that conditions can, do, and will change due to global warming. You can't rely on hurricane data from the last 50 years to predict the next 50. At best I hope they integrate this into the simulation models so that we don't have to brute force our calculations as much.
Re: (Score:2)
Read the article summary again.
Historical weather data fed into AI predictive software beat the "real scientists" 90% of the time with 99% better prediction of weather variables.
So relying on historical data CAN beat the current models. And beats it in nearly EVERY case!
Maybe someone needs to delete the AI software before we're lead to believe that the global warming proclamations are largely based on alarmism.
Re: (Score:2)
No it only outperformed 1 weather research center in Europe. I haven't read the paper, but I am skeptical of relying on machine learning for predictions in the long-term. It's also kind of disingenuous of them to claim that the ML process is less energy intensive just because it only took "minutes" to make a prediction. They fail to mention the amount of time and resources it took to train the ML model.
URL (Score:1)
https://charts.ecmwf.int/produ... [ecmwf.int]
I'll take this with a pinch of salt. (Score:2)