WeatherNext 2 on Windy.com ?!
-
Hi Windy.com Team
just watched the Google / DeepMind Spot on their new WeatherNext 2 AI Model.
Will, and if yes, when will we see it on Windy.com ?
Thanks!
-
Someone else may jump in to correct me....
The issue with AI models is that they usually (Haven't researched this AI model in particular), they rely heavily on past data rather than the normal NWP models which compute trillions of math equations to calculate the weather you see on windy.com
As I read on another forum earlier today. Someone pointed out that, with a expected SSW event to occur sometime at the end of the month. All the NWP models have predicted some variation of it while the AI models can't "see" it because the last SSW to occur in Nov was 1968 (I believe) so the past data is not there and so the AI models think that it is very rare as such will not happen.
I apologise if is a little critical (or wrong) but hope it gets my point across.
Let us know if you have anymore questions.
(P.S. I am not a Windy Team Member, so wait to see what they say)
-
@vhalx Agreed. As a subscriber, I think it would be awesome to get this state of the art model into Windy. Looking at their blog post, they seem to have much better forecasts than ECMWF:
https://blog.google/technology/google-deepmind/weathernext-2/
Living in a place with regular hurricanes, the resolution and error margin of models like ECMWF are pretty bad (both on the trajectory and wind speeds). -
@jeanr said in WeatherNext 2 on Windy.com ?!:
they seem to have much better forecasts than ECMWF:
What are you basing this statement off?
Read the article this morning, and as I expected, it uses Machine Learning.
-
W WeatherMan10 referenced this topic on
-
@WeatherMan10 you're right, I forgot to post the benchmark:
https://developers.google.com/weathernext/guides/evals
Note that using machine learning is akin to developing sophisticated forecasting heuristics, and it doesn't surprise me that they would handily beat traditional models like ECMWF, which have to solve extremely unstable PDEs, sometimes leading to unstable forecasts within 24 hours of a typhoon -
Hello, I will provide more info here after certain discussion in Windy team, which already started. Thank you for your patience.
-
I would also love to see the WearherNext 2 model get added to Windy!
-
@Suty do you have any update on this?
-
@salamoc Unfortunately, now, I don't have any further info about proceeding this new model.
-
@Suty it's a shame windy seems so slow to adopt new AI tools, not as a replacement but an addition to existing numerical forecasts
-
@jeanr Hello, we standardly evaluate these large AI models and we are careful about the implementation, since their data are not that accurate and also every model added to our app is an extra data that must be optimized and handled by the whole team. So we are only cautious about these right now.
-
@Suty this new version of Google's model is very accurate, especially for high impact events like tropical storms (where numerical models often fail). That's why having both numerical and AI models would be a great combination. Is it a problem of cost on your side ? Does Google charge a lot for live updates? Obviously models like GFS are free, but I think providing more accurate storm models would be great for premium users. Even the old GFS is moving away from pure numerical forecasts:
https://www.noaa.gov/news-release/noaa-deploys-new-generation-of-ai-driven-global-weather-models -
Another item that needs to be taken into account with a new model is Verification Stats meaning how accurate a weather model is at a certain step (time frame) in the forecast.
I had a look for the WeatherNext2 stats and I could not find them but I'll keep looking as I only had a quick look.
EDIT: Removed incorrect statement and chart
-
@WeatherMan10
This root mean square error (RMSE) graph illustrates the forecast error for the 500 hPa geopotential height. It does not mean that the model will provide the same accuracy at ground level for all parameters (wind, temperature, cloud cover, etc.). Furthermore, resolution is a crucial factor to consider for certain types of forecasts and locations.
According to Google, the resolution of WeatherNext 2 is 0.25° (23km)EDIT: the graph was removed by the user.
-
@idefix37 said in WeatherNext 2 on Windy.com ?!:
@WeatherMan10
This root mean square error (RMSE) graph illustrates the forecast error for the 500 hPa geopotential height. It does not mean that the model will provide the same accuracy at ground level for all parameters (wind, temperature, cloud cover, etc.). Furthermore, resolution is a crucial factor to consider for certain types of forecasts and locations.
According to Google, the resolution of WeatherNext 2 is 0.25° (23km)Was half asleep whilst writing that this morning then. 23km is really not good. I apologise about my wrong statement.
