Aardvark beats groundhogs and supercomputers in weather forecasting
(2025/03/26)
- Reference: 1742978109
- News link: https://www.theregister.co.uk/2025/03/26/aardvark_weather_forecast/
- Source link:
Aardvark, a novel machine learning-based weather prediction system, teases a future where supercomputers are optional for forecasting - but don't pull the plug just yet.
Academics affiliated with the Alan Turing Institute in the UK and other institutions claim they have developed a weather prediction model that can be trained and run on a desktop computer at a fraction of the cost and time currently incurred by using supercomputers.
Building on work done at the University of Cambridge, Richard Turner, research lead in AI for weather prediction at the Alan Turing Institute and a professor of machine learning at the University of Cambridge, and Scott Hosking, interim director of science and innovation at The Alan Turing Institute, [1]reckon Aardvark's results are good enough that it can replace the entire [2]numerical weather prediction (NWP) pipeline.
[3]
The NWP pipeline follows a three-step process. First, observational data is gathered from satellites, weather balloons, ground stations, ships, aircraft, and buoys, and combined with a recent forecast to estimate the current state of the atmosphere. Second, this estimate is fed into a complex computational model that simulates atmospheric physics to generate forecasts. Finally, those raw forecasts are post-processed to correct biases, improve local accuracy, and incorporate human forecaster input.
[4]
[5]
Other recent work improving weather forecasting with machine learning, such as Google DeepMind's [6]GenCast model , has focused on step two – the computational model – whereas Aardvark is said to be capable of replacing all three steps.
As described in [7]an article published in the science journal Nature this month, "Aardvark provides accurate forecasts that are orders of magnitude quicker to generate than existing systems, without any reliance on NWP products at deployment time. Generating a full forecast from observational data takes approximately one second on four NVIDIA A100 GPUs, compared to the approximately 1,000 node-hours required by [8]HRES to perform data assimilation and forecasting alone, before accounting for downstream local models and processing."
[9]
A quick video demo of the software is below.
[10]Youtube Video
The resulting forecast, it's claimed, is as accurate as America’s [11]Global Forecast System (GFS). Yet it relies on only about a tenth of the observational data used in traditional NWP systems.
[12]
"We only used 10 percent of the data as we’re quite limited in terms of computational resources in my academic lab compared to, say, technology companies," Turner told The Register . "With more data, I would expect the system to perform better. Aardvark is able to compete using only a fraction of the data partly because the AI models are trained directly to solve the task of interest (forecasting)."
[13]The secret to better weather forecasts may be a dash of AI
[14]NASA, IBM just open sourced an AI climate model so you can fine-tune your own
[15]Nvidia's latest AI climate model takes aim at severe weather
[16]Here's the ugliest global-warming chart you'll ever need to see
[17]Humans brought the heat. Earth says we pay the price
Aardvark still requires some refinement as it does not yet have the resolution of Europe's [18]Integrated Forecast System (IFS). That's being worked on, however, and the researchers expect to add various specialized modules to focus on specific types of forecasts like hurricanes, floods, and other extreme weather events.
Aardvark consists of three components: An encoder, a processor, and a set of decoder modules. The encoder has about 31 million parameters and takes 13 hours to train. The processor contains about 54 million parameters and requires eight hours of training on [19]ERA5 , a weather data set, followed by 3 hours of fine-tuning using the encoder's output. There are eleven decoder modules, each of which has about 2 million parameters and takes 30 minutes to train. Allowing a further two hours for end-to-end fine-tuning, Aardvark can be trained in about 100 GPU hours.
Once that's done, Aardvark can create forecasts on a desktop computer within minutes. As a point of comparison, Google claims its GenCast model can produce a 15-day forecast in eight minutes using a single Google Cloud TPU.
[20]Relevant source code to replicate the results in the Nature paper is presently restricted, though the plan is to eventually let Aardvark loose.
"Everything will be open sourced when the print version of Aardvark goes live and anyone will be able to download it and train the system on tasks of interest," said Turner. ®
Get our [21]Tech Resources
[1] https://www.turing.ac.uk/blog/project-aardvark-reimagining-ai-weather-prediction
[2] https://www.ncei.noaa.gov/products/weather-climate-models/numerical-weather-prediction
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Z-PeVq1OlDU_Amfm2JWoIwAAAJU&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z-PeVq1OlDU_Amfm2JWoIwAAAJU&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z-PeVq1OlDU_Amfm2JWoIwAAAJU&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://www.theregister.com/2024/12/05/google_deepmind_weather_model/
[7] https://www.nature.com/articles/s41586-025-08897-0
[8] https://climalert-docs.imida.es/en/docs/data/nwp/ifs.html
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z-PeVq1OlDU_Amfm2JWoIwAAAJU&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://youtu.be/zAovQbs6jlw?feature=shared
[11] https://www.ncei.noaa.gov/products/weather-climate-models/global-forecast
[12] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z-PeVq1OlDU_Amfm2JWoIwAAAJU&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[13] https://www.theregister.com/2024/07/27/google_ai_weather/
[14] https://www.theregister.com/2024/09/25/nasa_ibm_ai_weather/
[15] https://www.theregister.com/2024/08/19/nvidia_ai_weather/
[16] https://www.theregister.com/2025/02/23/ugliest_global_warming_chart/
[17] https://www.theregister.com/2025/02/02/heatwaves_future/
[18] https://www.ecmwf.int/en/forecasts/documentation-and-support
[19] https://www.ecmwf.int/en/forecasts/dataset/ecmwf-reanalysis-v5
[20] https://zenodo.org/records/13158382
[21] https://whitepapers.theregister.com/
Academics affiliated with the Alan Turing Institute in the UK and other institutions claim they have developed a weather prediction model that can be trained and run on a desktop computer at a fraction of the cost and time currently incurred by using supercomputers.
Building on work done at the University of Cambridge, Richard Turner, research lead in AI for weather prediction at the Alan Turing Institute and a professor of machine learning at the University of Cambridge, and Scott Hosking, interim director of science and innovation at The Alan Turing Institute, [1]reckon Aardvark's results are good enough that it can replace the entire [2]numerical weather prediction (NWP) pipeline.
[3]
The NWP pipeline follows a three-step process. First, observational data is gathered from satellites, weather balloons, ground stations, ships, aircraft, and buoys, and combined with a recent forecast to estimate the current state of the atmosphere. Second, this estimate is fed into a complex computational model that simulates atmospheric physics to generate forecasts. Finally, those raw forecasts are post-processed to correct biases, improve local accuracy, and incorporate human forecaster input.
[4]
[5]
Other recent work improving weather forecasting with machine learning, such as Google DeepMind's [6]GenCast model , has focused on step two – the computational model – whereas Aardvark is said to be capable of replacing all three steps.
As described in [7]an article published in the science journal Nature this month, "Aardvark provides accurate forecasts that are orders of magnitude quicker to generate than existing systems, without any reliance on NWP products at deployment time. Generating a full forecast from observational data takes approximately one second on four NVIDIA A100 GPUs, compared to the approximately 1,000 node-hours required by [8]HRES to perform data assimilation and forecasting alone, before accounting for downstream local models and processing."
[9]
A quick video demo of the software is below.
[10]Youtube Video
The resulting forecast, it's claimed, is as accurate as America’s [11]Global Forecast System (GFS). Yet it relies on only about a tenth of the observational data used in traditional NWP systems.
[12]
"We only used 10 percent of the data as we’re quite limited in terms of computational resources in my academic lab compared to, say, technology companies," Turner told The Register . "With more data, I would expect the system to perform better. Aardvark is able to compete using only a fraction of the data partly because the AI models are trained directly to solve the task of interest (forecasting)."
[13]The secret to better weather forecasts may be a dash of AI
[14]NASA, IBM just open sourced an AI climate model so you can fine-tune your own
[15]Nvidia's latest AI climate model takes aim at severe weather
[16]Here's the ugliest global-warming chart you'll ever need to see
[17]Humans brought the heat. Earth says we pay the price
Aardvark still requires some refinement as it does not yet have the resolution of Europe's [18]Integrated Forecast System (IFS). That's being worked on, however, and the researchers expect to add various specialized modules to focus on specific types of forecasts like hurricanes, floods, and other extreme weather events.
Aardvark consists of three components: An encoder, a processor, and a set of decoder modules. The encoder has about 31 million parameters and takes 13 hours to train. The processor contains about 54 million parameters and requires eight hours of training on [19]ERA5 , a weather data set, followed by 3 hours of fine-tuning using the encoder's output. There are eleven decoder modules, each of which has about 2 million parameters and takes 30 minutes to train. Allowing a further two hours for end-to-end fine-tuning, Aardvark can be trained in about 100 GPU hours.
Once that's done, Aardvark can create forecasts on a desktop computer within minutes. As a point of comparison, Google claims its GenCast model can produce a 15-day forecast in eight minutes using a single Google Cloud TPU.
[20]Relevant source code to replicate the results in the Nature paper is presently restricted, though the plan is to eventually let Aardvark loose.
"Everything will be open sourced when the print version of Aardvark goes live and anyone will be able to download it and train the system on tasks of interest," said Turner. ®
Get our [21]Tech Resources
[1] https://www.turing.ac.uk/blog/project-aardvark-reimagining-ai-weather-prediction
[2] https://www.ncei.noaa.gov/products/weather-climate-models/numerical-weather-prediction
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Z-PeVq1OlDU_Amfm2JWoIwAAAJU&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z-PeVq1OlDU_Amfm2JWoIwAAAJU&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z-PeVq1OlDU_Amfm2JWoIwAAAJU&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://www.theregister.com/2024/12/05/google_deepmind_weather_model/
[7] https://www.nature.com/articles/s41586-025-08897-0
[8] https://climalert-docs.imida.es/en/docs/data/nwp/ifs.html
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z-PeVq1OlDU_Amfm2JWoIwAAAJU&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://youtu.be/zAovQbs6jlw?feature=shared
[11] https://www.ncei.noaa.gov/products/weather-climate-models/global-forecast
[12] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z-PeVq1OlDU_Amfm2JWoIwAAAJU&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[13] https://www.theregister.com/2024/07/27/google_ai_weather/
[14] https://www.theregister.com/2024/09/25/nasa_ibm_ai_weather/
[15] https://www.theregister.com/2024/08/19/nvidia_ai_weather/
[16] https://www.theregister.com/2025/02/23/ugliest_global_warming_chart/
[17] https://www.theregister.com/2025/02/02/heatwaves_future/
[18] https://www.ecmwf.int/en/forecasts/documentation-and-support
[19] https://www.ecmwf.int/en/forecasts/dataset/ecmwf-reanalysis-v5
[20] https://zenodo.org/records/13158382
[21] https://whitepapers.theregister.com/
I'm happy as long as everyone keeps using different models, I'm happy.
Anonymous Coward
I know deep down I shouldn't do this, but I usually look at the weather forecast from several sources and decide which I like the best.
If it ain't broke, don't fix it.
Yorick Hunt
I still prefer the weather stone, been accurate since its creation;
https://en.wikipedia.org/wiki/Weather_rock
Re: If it ain't broke, don't fix it.
jake
That's not forecasting, that's a form of local nowcasting.
Less is more
Charlie Clark
Can't read the paper, but this looks like another example that picking the right parameters is key to any statistical exercise, which underpin most of modern AI systems. Starting from relatively simple machine learning, LLMs just added parameters as they came along in the race to be "first". Having now more or less reached the "complexity horizon", other approaches are required.
I use the Golden Gate Bridge.
jake
I'm only accurate out about four days, but some old-timers get pretty good accuracy out almost a week.
It works by using the bridge as a kind of barometer. The fog hits the bridge at different heights depending on barometric highs and lows, Their relative movements can be visually tracked, both how far off the water, and how much of the towers can be seen above it. Couple that with observing the wind direction, if one tower is clear and the other is in fog (changing wind directions) and various other factors combine to make for a very useful tool to the observant local. Knowing what the water temp readings readings are at the off-shore buoys are help, too, but some say that's cheating.
Re: I use the Golden Gate Bridge.
Anonymous Coward
I'm surprised that gives accurate forecasts for the UK regions but I trust people on the internet.
Simple model
Weather(tomorrow) = Weather(today)
Accurate about 80% of the time, but sometimes gets it drastically wrong.