AI set to revolutionise weather forecasting?

[Derecho]

Very interesting article this. I'll put the key pieces of text in below for those who want a more simplified outline:
 

 

Just a moment...

WWW.SCIENCE.ORG

Today, the 6-day forecast is about as good as the 3-day forecast from 30 years ago. It also comes with a cost: billions of dollars’ worth of energy-hungry supercomputers that must run 24/7 just to produce a few forecasts a day.

In mere minutes on cheap desktop computers, trained AI systems can now make 10-day forecasts that are as good as the best traditional models—and in some cases even better.  The algorithms could enable more frequent forecasts and free up computing resources.

Traditional weather models start by feeding a snapshot of current conditions, based on observations from satellites, weather stations, and buoys, into a gridlike computer model that divides the atmosphere into millions of boxes. The snapshot is run forward in time by applying the physical laws of fluid dynamics to each box.

The new AI models skip the expense of solving equations in favor of “deep learning.” They identify patterns in the way the atmosphere naturally evolves, after training on 40 years of ECMWF “reanalysis” data—a combination of observations and short-term model forecasts that represents modelers’ best and most complete picture of past weather.

When fed a starting snapshot of the atmosphere based on the same combination of observations and modeling, GraphCast can outperform the ECMWF forecast out to 10 days on 90% of its verification targets, including hurricane tracks and extreme temperatures.

To improve further, the AI models could be weaned off the reanalysis data, which carry the biases of traditional models. Instead, they could learn directly from the petabytes of raw observation data held by weather agencies, Keisler says. Google’s short-term weather model already does so, training itself on data from weather stations, radar, and satellites.

However, note the final paragraph:

Adoption might be slowed by unease about the black-box nature of the AI: Researchers often can’t say how such systems reach their conclusions.

So instead of relying on physical theory to produce weather models, AI is taking past reanalysis models to train data. This can be run several times a day and is currently comparable to EC model output and in some cases better.

When we start assimilating even more observations from the past we should see even more improvements... but how good can AI forecasting become?

Edited January 29 by Derecho

[Relativistic]

  @Derecho

2 hours ago, Derecho said:

Adoption might be slowed by unease about the black-box nature of the AI: Researchers often can’t say how such systems reach their conclusions.

I see this as an amazing opportunity to learn. These AI models have developed a strong intuition for weather forecasting and delving under the blackbox to see how they're able to pick out the right patterns should yield all sorts of insights.

Something to embrace IMO.

Edit: Worth adding that this applies to any field where AIs are outperforming previous technologies.

Edited January 29 by Relativistic

[Derecho]

  @Relativistic Indeed it does but given the complex nature of the current models, there is a black box element to them anyway.

Either way I think there will be big advances in forecasting over the next decade. For example, how much better will 40 years worth of observations / analysis data going into a model compare to theoretical assumptions that are currently used over the Greenland region?

 

Edited January 29 by Derecho

[Paul]

Some extra info and interesting reading on this subject here too:

GraphCast: AI model for faster and more accurate global weather forecasting

DEEPMIND.GOOGLE

---

Our state-of-the-art model delivers 10-day weather predictions at unprecedented accuracy in under one minute

WeatherBench: Weather Forecasting using ML - Google Research

SITES.RESEARCH.GOOGLE

---

Discover how Google's WeatherBench is providing users with an updated weather forecasting benchmark using machine learning.

 

[lassie23]

AI can't even write an article without copying and filling the word count with waffle, so forecasting the weather is a step too far at the moment lol

[Earthshine]

 lassie23 I wouldn't be so sure, the AIFS has recently been updated by the ECMWF and has skill just as good as the physics-based IFS using attention graph neural networks: https://www.ecmwf.int/en/about/media-centre/aifs-blog/2024/first-update-aifs

[The PIT]

Interesting this suggests the forecast is predictal and not chaotic.

[Earthshine]

 The PIT That's actually a problem with "AI" models.  They are trained to minimise the absolute error between the output and the training reanalysis dataset.  When used operationally, the output that is an optimum solution is a "blurred" forecast with no extremes at longer lead times.  You can actually see the kinetic energy just randomly drop away.  I've applied for a grant to work with a software engineer to hopefully incorporate physical constraints into the training of these models.  Hopefully then the output will be constrained to not only fit the observational data, but to actually force it to conserve kinetic energy, mass, etc.  These data-driven models are the very first generation of operational models.  I would be EXTREMELY hesitant to say that they will never overtake classic physics based models.  The rate of progress has just exploded - and these are basically equivalent to the very first generation of classic NWP models.  Personally I think the biggest constraints to better weather forecasts at the moment is a lack of observations - satellites are great, but the vertical resolution of sounders is seriously lacking.

[Relativistic]

 The PIT Chaos is a property of physical systems; the weather will always be chaotic.

 Earthshine That's interesting about kinetic energy dropping away given that they've been trained on datasets that do conserve energy. One would expect fluctuations but a specific bias is somewhat surprising.

Good luck with the grant application!

[Mike Poole]

Really interesting to see how AI forecasting develops.  I’ve been impressed by what I’ve seen from the ones on the ECM website, astounding, actually, that they can even hold their own with the big NWP models.

Presumably though, none of the pressure fields that they generate are actually constrained by the laws of physics?  For the NWP, since they simulate the laws of physics (or their best approximation) all patterns generated are physically possible - that wouldn’t be the case with AI.  Does this matter?  If it gets the gist right, would people care?

There must be a huge amount that AI could learn from the reanalysis data - but we wouldn’t get the understanding, just the output, who knows what teleconnections AI might discover but have no way of communicating to us.  Fascinating new story of weather models, which we are only just at the beginning of.

[Earthshine]

 Mike Poole You're exactly right.  There is no inherent knowledge of physics, only inferred from the ERA5 training data.  I'm amazed it works, it seems such an ill-posed problem yet here we are.  On the problem of not respecting physical laws, this is a problem because the atmosphere is chaotic and if you aren't conserving energy the intensity of storms is underestimated.  We saw this with AI forecasts for Storm Ciaran which were weaker than the physics-based IFS forecast.

[Earthshine]

 Relativistic Thanks!  It's an interesting problem.  Really the model is trained to minimise the difference in variables between the ERA5 dataset and it's own output (Pangu is four surface and 5 upper air variables - MSLP, T u, v winds + mixing ratio for upper air variables on 13 pressure levels).    It infers the physics on how these variables interact with zero knowledge of the known physics, something I find extraordinary that it works!

[Derecho]

 Earthshine Seems like reliance on reanalysis models constrains the output of AI models then. Maybe tweak them to apply the laws of physics in areas where we have a deepening low only? Leave the rest to AI?

The interesting thing is... given the AI performs just as good and potentially better then the physics driven models, there must be aspects of physics the AI model is incorporating that we don't understand yet?

[Earthshine]

 Derecho A lot of NWP is parameterised where the scale of events is just too small to capture in the model (e.g. really local convective events).  In theory AI isn't really limited to a grid, I've seen neural fields allow variables to be mapped to arbitrary resolution (e.g. your neural network maps x, y, z coordinates f(x,y, z) --> t,p,q with no reference to an actual grid).  Probably won't work in NWP, but we already can already see upscaling of precipitation using AI show encouraging results: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2022MS003120

[Derecho]

 Earthshine Interesting. From working with reanalysis models in the past, I know they are able to go far back in time because they assimilate little in the way of variables. The 20CR for example goes back to the early 1800s because it only assimilates SLP, SST and to an extent sea ice into a weather model.

I'm guessing that running them in the level of detail to capture small scale events like local convection just isn't possible with the number of variables in earlier years and computing power available? So are other methods are required to capture this detail and still being worked upon?

[Rain All Night]

Anyone else looking forward to the presumably inevitable machine learning model ensemble suites becoming available?

[Mike Poole]

 Rain All Night If each run takes 1 minute, 1000 member suites are easily possible…

[MattStoke]

Hopefully one day we'll also have quantum computers producing forecasts in real time.

[LetItSnow!]

 Mike Poole The forums will be on overload in five years if that's the case! I can picture it now... 996 mild looking charts and people posting the four that show a beasterly in 12 days time!

 

In all seriousness, I think AI could potentially be a major shift in weather modelling and I can't wait to see the developments in the future. A super smart AI may be able to pick out hard to predict features that currently only pop up at short notice. Also, with a changing atmosphere, AI may have a better tackling at the effects that has.

[Methuselah]

Maybe fungi will play an important part in AI? 

 

[lassie23]

The AI met office, weekly deep dive,

[Paul]

An interesting study looking at the major AI models Vs the traditional ones during storm Ciaran. 

Do AI models produce better weather forecasts than physics-based models? A quantitative evaluation case study of Storm Ciarán

WWW.NATURE.COM

---

npj Climate and Atmospheric Science - Do AI models produce better weather forecasts than physics-based models? A quantitative evaluation case study of Storm Ciarán

 

 

[Blessed Weather]

I have started a new section in the Netweather Research Library with articles, blogs and research papers covering the use of AI in weather modelling and forecasting. I've included a couple of links to some interesting 'learning' material written/presented by Harvard and Cambridge Universities.

If anyone has any more research papers or articles they believe should be included in the library, please either post them in here (and tag me) or message me. Thanks.