12 ms·
MetNet-3: A state-of-the-art neural weather model
- photochemsyn 3y agoInteresting capabilities but why don't they report the capabilities of the model beyond a 24-hour cutoff? What do their 5-day forecasts look like? [edit] Modern numerical prediction models are pretty good in the five-day range (~90%), I'm guessing the deep learning models diverge rapidly in comparison (though perhaps they're better in the sub-24 hour range). Both approaches benefit from more extensive data collection systems as inputs. See (full text): "Advances in weather prediction" Alley et. al Science 2019 https://par.nsf.gov/servlets/purl/10109891 https://par.nsf.gov/servlets/purl/10109891
- vore 3y agoDoes anyone's 5 day forecast look good?
- codeslave13 3y ago5 day is closed to wild ass guess in my area. Mountains complicate everything
- orangepurple 3y agoAnything over 24 hours is pretty much always very wrong in the mid Atlantic region.
- capableweb 3y agoHow often do you need weather services in the middle of the Atlantic though? Seems like a pretty niche use case, except for the various islands there. Edit: I thought something was sketchy and rightly so, after searching for "Mid Atlantic Region" I learned that it's actually a region in the north-east US, not "the middle of the Atlantic". Well, learned something new today :)
- vore 3y agoMid Atlantic = south of New England north of Virginia (more or less), not middle of the Atlantic Ocean.
- hgomersall 3y agoNot to be confused with a mid-Atlantic accent, which is possibly what you might get around the middle of the Atlantic if your interpolated a couple of specific local accents between the UK and the US
- AlotOfReading 3y agoThe Mid-Atlantic [coast] is part of the US east coast around NY, NJ, Pennsylvania, etc rather than literally the middle of the Atlantic Ocean. It's a fairly important region considering how much of the US and global economy reside there.
- Workaccount2 3y agoAs someone who both lives in the mid-atlantic and regularly has to plan around the weather, I can assure you that you simply have a confirmation bias. Our forecasts are actually pretty good, even for ones 7 days out. Where people tend to get thrown is micro-storms during the summer months. They are basically impossible to predict accurately, at best it's just known that a random assortment of towns in a given area will received heavy rain for a short time. Being able to read radar is the best way to deal with this, but it's very short term only (15min to 1 hr).
- orangepurple 3y agoReading radar is even harder because you need a 3D view of the atmosphere to understand why storms are coming seemingly out of thin air. This is especially a problem near rivers. We only get a 2D doppler slice.
- Workaccount2 3y agoYou just have to look at the 2D slice and see if a storm pocket is headed your way. That's the best its gonna get for a layman.
- X6S1x6Okd1st 3y agoyes? https://charts.ecmwf.int/products/plwww_m_hr_ccaf_adrian_ts?player_dimension=single_product&single_product=latest https://charts.ecmwf.int/products/plwww_m_hr_ccaf_adrian_ts?...
- bglazer 3y agoI get the impression this model is meant for very localized, short term prediction. Like whether or not its going to be raining in your neighborhood in the next hour.
- lainga 3y agoIt looks to me like it's "upscaling" ENS data to high resolution. Global general circulation models work with somewhat low-resolution data about terrain, and this model has found a function mapping the low-resolution weather predictions back to a prediction on high-resolution terrain. (ed.: true also, but to a lesser extent, for "mesoscale" models (e.g. of just North America with boundary conditions to a global model)) If it did learn longer-range predictions (or the next model does?), I would hazard the model had achieved speedup by internalising the patterns of certain large-scale weather connections, e.g. the jet streams, Walker circulation, ENSO, Gulf stream... which I think will be fine for 99% of cases, the 1% being if these established patterns break somehow. ("freak weather") At that point you would have to return to a general circulation model. When you take away the long-lived circulatory features that are familiar to us, and that are particular to Earth, predicting the weather is "just" fluid dynamics. These are both just wild guesses, though
- xrd 3y agoThis is interesting in that I've noticed when a hurricane is about to hit Florida you have to filter through all this spam news to get actual actionable information about the storm. The terrible news agencies have no incentive to provide the information because they just need to provide click baity links to serve ads, and it feels like the information you eventually get isn't the most accurate, it's the most sensational. "THIS might be the biggest storm in thousands of years! PLEASE SHARE THIS LINK ON SOCIAL MEDIA!" If this model can be used by independent media or by me, I could provide a blog which gives accurate information and actually helps people. That's a very interesting turn. I can't tell if this model is released publicly from this article or just available behind a Google service? And, if it helps further the demise of these consolidated "local" news sites (which are always just content mills owned by some large national owner) then even better.
- jszymborski 3y agoIn Canada, I use weather.gc.ca . It's steered me through some terrible ice storms. https://www.weather.gc.ca/city/pages/qc-147_metric_e.html https://www.weather.gc.ca/city/pages/qc-147_metric_e.html Does weather.gov suit your purposes? https://www.weather.gov/mfl/ https://www.weather.gov/mfl/
- switchbak 3y agoI’ve found the Canadian weather site to be quite inaccurate as compared to a local forecast from weather underground. Especially when it comes to short term predictions of actual rain, it seems magnitudes better, and it updates its forecasts at a much higher frequency. The precipitation view is only available on the web view for some reason. That said, the Weather Canada satellite view is indispensable. Even if the site hasn’t changed in literally 25 years.
- bee_rider 3y agoWhat actions are necessary on your part? Having been lucky enough to grow up in New England, my response to cold weather stuff is mostly… go inside, get a blanket and throw another log on the fire. But in Canada you all get a more serious type of cold I think. Places like Florida or Kansas where the weather will actually come get you inside seem like pretty out there places to live.
- andy_xor_andrew 3y agoHold on, Google has its own weather model? Are they they only one that use it, or do agencies use them as well?
- mwest217 3y agoI wish this had an API to be used via other weather apps. There are many native iOS weather apps whose interface I prefer to the Google Search app, and I'm sure the big ones (like Carrot Weather) would add this as a weather source if it were an option.
- baq 3y agoI'd love to see how it fared with last month's Otis since it caught basically all the currently used models off guard.
- m3kw9 3y agoProblem with these models coild be that they are trained with historical weather and patterns, but when new weather effects come up they will get worse over time.
- NavinF 3y agoThese models are not as simplistic as you imagine.
- dekhn 3y agothey might not be simplistic, but we've already got experience with google making a model and then external conditions change, invalidating the model- see Google Flu Trends. After the team launched it and got their promos and moved on, the zombie jobs training the model failed to make useful predictions the next year. The model was not simplistic; it's just that it didn't generalize and needed constant attention from humans.
- m3kw9 3y agoMaybe not as magical as you imagine
- potatoman22 3y agoYou're completely right. This model, as with any other predictive model, is subject to degredation in performance when the data-generating process changes. But given MetNet-2 came out in 2021, they'll probably release an updated version before the performance degrades due to changes in weather patterns.
- bitshiftfaced 3y agoOpposed to training them on future weather and patterns?
- waveBidder 3y agoopposed to first principles modeling. the problem isn't that we don't have good weather models, the problem is they're chaotic, so you need exponentially more data for linear increases in forecast window. and I do literally mean exponential there.
- ppaattrriicckk 3y agoIt might not come as much of a shocker, but a couple of researchers last year claimed a strong correlation between raw compute and predictive power in various fields/domains, including weather forecasting: https://arxiv.org/pdf/2206.14007.pdf#page=10 https://arxiv.org/pdf/2206.14007.pdf#page=10 (page 10 specifically has graphs on Weather Forecasting vs. Compute). In their study they claimed a strong correlation in these fields (vs. compute): * Weather Forecasting * Protein Folding * Oil Exploration (at BP) * Chess * Go ... The latter 2 being games, which I personally do not find surprising. But I do find it inspiring that we can "just" calculate our way out of some important issues. That hopefully translates well to other fields.
- pkdpic 3y agoThank you for that link, Ive been subconsciously holding off on assuming there was a compute / predictive power correlation even though it seems natural. But it would probably be dangerously naïve to have assumed that connection. Anyway good for us! Go humans! (computers)
- ppaattrriicckk 3y agoGlad you found it useful. A small caveat, though: The correlation is linear with the logarithm of compute. So here's hoping Moore's law & friends live on a tad longer! And a somewhat unrelated fun fact: The authors surprisingly found the lowest correlation between compute and the performance in the domain of Go (and not the real world). Although the data is very sparse, I suspect that it's due to algorithmic advances.
- p_j_w 3y ago>I do find it inspiring that we can "just" calculate our way out of some important issues. In the case of oil exploration, we can calculate our way into some!
- seabass-labrax 3y agoCalculate, suffocate, annihilate. The fact that we now need computers to find oil says a lot about how much we've already used up, and the precarious position we find ourselves in as a result.
- fallat 3y agoWhere can I pay for an API that I can use personally? Seriously, I'm sold. I've seen other prediction models. This one looks fantastic.
- 1970-01-01 3y agoThe real surprise here is how well the product resembles the early Google mindset. A surprise release of something very useful, working far better than the competition, and also free to use.
- _ea1k 3y agoYeah, it is as if some group forgot the new "Be Evil" motto.
- ctoth 3y agoMy silly question for the day: Could you use interpretability techniques on this model to figure out the easiest places to perturb conditions to steer towards particular weather in a particular location (cloud seeding, or whatever?) Is this the first part of the weather control machinery from Star Trek? In order to control, one must first predict?
- counters 3y agoIt doesn't seem like MetNet outputs a complete 3D atmospheric state, just specific (and mostly surface-level) forecast predictands [1]. The analysis you're describing could be done with a traditional numerical weather prediction and data assimilation system (especially if that system implements 4DVar, since you'll already have the model adjoint available - well, technically the tangent linear, but still applicable here). [1]: https://arxiv.org/pdf/2306.06079.pdf https://arxiv.org/pdf/2306.06079.pdf
- loxias 3y agoI love this line of thought, I've also entertained it from time to time in other domains. (macroecon :D) To some extent, yes, but you'd need more energy than is practical. Weather is a chaotic system -- future behavior can be highly sensitive to local fluctuations.
- SomewhatLikely 3y agoThat input might be too far outside of the training conditions for it to believe the sensors. It would probably assume those sensors you perturbed are faulty and ignore them.
- kposehn 3y agoI was just recently wondering when we'd see some new weather models released that keep pace with the development of machine learning. Now if I could just access the raw forecast data from MetNet-3...
- zuzun 3y agoIt's already happening. For example, ECMWF provides experimental forecasts by their own deep learning model, plus models from Deepmind, NVIDIA, and Huawai.
- ckrapu 3y agoCool work and I hope to be able to use it some day. Most importantly though, does anyone know how they made the animation with the data sources? I feel like that came from something lightweight and convenient and I'd like to know what it was.
- cryptoz 3y agoAll this research from Google and still no discussion of using their access to billions of barometers from Android devices? sigh.
- spdustin 3y agoHere's the paper about MetNet-3: https://arxiv.org/abs/2306.06079 https://arxiv.org/abs/2306.06079
- the_doctah 3y agoWhen I search with my Google app it says "Source: weather.com"
- YossarianFrPrez 3y agoHow cool! The paper was last revised on Arxiv over the summer; this blog post announces that MetNet3 is now powering weather predictions across google devices and services. (I'll bet it gets picked up and used for electricity demand forecasting if it hasn't already.) From the paper: > While ground based radars provide dense precipitation measurements, observations that MetNet-3 uses for the other variables come from just 942 points that correspond to weather stations spread out across Continental United Stated (CONUS). I don't know a thing about weather prediction, but the fact MetNet-3 can do it using data from less than 1000 points across the continental US is surprising. The other line that stood out to me was: > On a high level, MetNet-3 neural network consists of three parts: topographical embeddings, U-Net backbone and a MaxVit transformer for capturing long-range interactions. If I understand it correctly, MetNet-3 is sort of abstractly treating 'predicting the weather at each geographical patch' like a very big computer vision problem.
- nelox 3y agoHow well did it forecast the rapid intensification of Hurricane Otis where other models failed?