Google MetNet - A Reliable Precip Model?

[Randy Jennings]

In s blog post and accompanying paper, researchers at Google detail an AI system — MetNet — that can predict precipitation up to eight hours into the future. They say that it outperforms the current state-of-the-art physics model in use by the U.S. National Oceanic and Atmospheric Administration (NOAA) and that it makes a prediction over the entire U.S. in seconds as opposed to an hour.

Google details MetNet, an AI model better than NOAA at predicting precipitation

Google says that its MetNet machine learning algorithm can outperform NOAA's precipitation forecasting model for timescales up to 8 hours ahead.

venturebeat.com

A Neural Weather Model for Eight-Hour Precipitation Forecasting

Posted by Nal Kalchbrenner and Casper Sønderby, Research Scientists, Google Research, Amsterdam Predicting weather from minutes to weeks ahead wi...

ai.googleblog.com

 

[rdale]

They are taking existing precip and moving it. That may be effective - but it doesn't account for precip that forms after FH0. So in essence - might be good for future project, but useless as-is.

 

[Mark Blue]

Anytime corporate America gets involved with a large-scale weather project I wouldn’t rest on the laurels of NOAA and NWS. With the likes of Google and their deep pockets I’d be looking over my shoulder for the day when privatization comes knocking at the front door. They have a massive amount of computational resources and can also pay salaries well beyond the standard GS11-13 meteorologists make in civil service after earning a PhD. I think it bears watching simply because we cannot seem to equal or best the Euro models. We’re closing the gap but still have work to do. Please don’t take offense @Jeff Duda as I know this is personal to you. After all it’s just my humble opinion.

 

[rdale]

I don't think anyone is suggesting otherwise But NWS short-range modeling is far superior to the ECMWF so make sure to apples and oranges...

 

[Jeff Duda]

Yeah, we're aware of it at ESRL/GSD (i.e., the HRRR group). Definitely some competition there. I haven't fully scrutinized the work yet (they didn't give adequate details to perform a rigorous scientific comparison anyway).

There is no doubt that machine learning/AI has a place in weather forecasting/NWP, and it certainly appears that, in this restricted framework, Google's software may indeed outperform the HRRR at very short ranges. One big thing that separates ML/AI forecasting from full-blown dynamical forecasting is that ML/AI is still restricted to past data sets and mostly serves as post-processing to existing forecast data, whereas the dynamical method actually uses the laws of physics directly to perform a weather forecast. Thus, ML/AI has an upper bound to how good it can forecast, especially temporally.