[HN Gopher] MetNet-3: A state-of-the-art neural weather model
___________________________________________________________________
MetNet-3: A state-of-the-art neural weather model
Author : apsec112
Score : 131 points
Date : 2023-11-03 00:34 UTC (22 hours ago)
(HTM) web link (blog.research.google)
(TXT) w3m dump (blog.research.google)
| photochemsyn wrote:
| Interesting 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
| vore wrote:
| Does anyone's 5 day forecast look good?
| codeslave13 wrote:
| 5 day is closed to wild ass guess in my area. Mountains
| complicate everything
| orangepurple wrote:
| Anything over 24 hours is pretty much always very wrong in
| the mid Atlantic region.
| capableweb wrote:
| How 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 wrote:
| Mid Atlantic = south of New England north of Virginia
| (more or less), not middle of the Atlantic Ocean.
| hgomersall wrote:
| Not 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 wrote:
| The 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 wrote:
| As 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 wrote:
| Reading 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 wrote:
| You 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 wrote:
| yes? https://charts.ecmwf.int/products/plwww_m_hr_ccaf_adrian
| _ts?...
| bglazer wrote:
| I 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 wrote:
| It 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 wrote:
| This 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 wrote:
| In 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
|
| Does weather.gov suit your purposes?
|
| https://www.weather.gov/mfl/
| switchbak wrote:
| I'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 wrote:
| What 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.
| foobarchu wrote:
| > Places like Florida or Kansas
|
| It's not too complicated with tornados. If you're urban,
| you listen for sirens, then take cover if you have to. If
| you have a basement, go there.
|
| Otherwise, you just watch the sky when it gets spooky and
| kind of accept you might get Oz'd at any time. There's not
| many prep actions to take, other than maybe popping open
| the garage door and getting out lawn chairs if you have a
| good view.
| jokethrowaway wrote:
| Having lived in Europe all my life, it sounds pretty
| insane to live like this.
|
| I'm thinking of moving to the Caribbean but storms /
| hurricanes are quite a mental jump for me.
|
| I guess that's the price to pay for hot water in the sea
| foobarchu wrote:
| Hurricanes are a different story, they are highly
| predictable (even in the very rare cases that the models
| are horrifically wrong about intensity, like the recent
| Hurricane Otis, we still got the track pretty close), but
| gigantic.
|
| Tornados on the other hand are almost impossible to
| predict. The best our weather service can do is say "this
| storm is the kind that produces tornados, watch out" (a
| tornado watch), and to set off the sirens when one is
| sighted (a tornado warning).
|
| Hurricanes are high intensity over a very large area,
| lasting for a long time. Tornados are short lived,
| unbelievably powerful, and cut a narrow path through
| whatever they decide to mow over.
| zombielinux wrote:
| The site(s) you are looking for are:
|
| nhc.noaa.gov and https://spaghettimodels.com/
| dudleypippin wrote:
| Similarly: https://www.trackthetropics.com/
| Quinner wrote:
| The vast majority of US weather data and forecasting comes from
| the National Weather Service, and you can access it directly at
| www.weather.gov
| baq wrote:
| There are only 3 places you need:
|
| 1. https://www.nhc.noaa.gov/
|
| 2. Your local NWS site
|
| 3. https://www.tropicaltidbits.com/
| rjj wrote:
| Simple. Find your dot on about 10 PDFs, interpret a handful
| of weather variables, know the safety tipping points of each,
| don't get it wrong or you may be injured, and check back
| every 4 hours! Easy.
| Workaccount2 wrote:
| Usually this is what most people would be looking for:
|
| https://www.tropicaltidbits.com/analysis/models/?model=ecmw
| f...
| jsight wrote:
| TBH, the NHC one is very good. Each storm has a "Forecast
| discussion" link with specific details on the things that
| specifically drove the forecast. The NWS publishes
| something similar for each area forecast, and it is often
| incredibly insightful.
|
| It isn't necessarily as good as the best local weather
| coverage, but it might help to point you to which station
| is giving the best coverage.
| spdustin wrote:
| And honestly, the local weather office forecast
| discussions are great, too. If they seem too dense with
| arcane language, that's actually something that ChatGPT
| does a great job of distilling. "Act like a professional
| meteorologist specializing in public speaking. Please
| read the following technical forecast discussion from
| NWS, and rephrase it to be more accessible to an audience
| that is educated, but not experts in meteorology."
| auspiv wrote:
| Do you want sensational click-bait articles or do you want
| actual weather forecasts by people who understand the topic
| and know how to interpret the models? Take your pick. One
| is simple, the other is not.
| andy_xor_andrew wrote:
| Hold on, Google has its own weather model? Are they they only one
| that use it, or do agencies use them as well?
| mwest217 wrote:
| I 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 wrote:
| I'd love to see how it fared with last month's Otis since it
| caught basically all the currently used models off guard.
| m3kw9 wrote:
| Problem 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 wrote:
| These models are not as simplistic as you imagine.
| dekhn wrote:
| they 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 wrote:
| Maybe not as magical as you imagine
| potatoman22 wrote:
| You'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 wrote:
| Opposed to training them on future weather and patterns?
| ppaattrriicckk wrote:
| It 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
| (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 wrote:
| Thank 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 naive to have assumed that connection. Anyway good
| for us! Go humans! (computers)
| ppaattrriicckk wrote:
| Glad 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 wrote:
| >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 wrote:
| Calculate, 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.
| auspiv wrote:
| Computers have been used in the industry since at least the
| 1970s. These days, every major oil company has $XXX million
| in supercomputers cranking through seismic data in a number
| of extremely computationally intensive ways. Reservoir
| simulation is also computationally intensive, but can be
| done at the individual workstation level.
|
| Here's a paper from May 1985 titled "Applications of
| supercomputers in the petroleum industry" - https://journal
| s.sagepub.com/doi/abs/10.1177/003754978504400.... Found
| that without even looking for oldest example.
| moffkalast wrote:
| Correct me if I'm wrong but given that weather patterns are
| fundamentally chaotic, at some point throwing more compute at
| the wall probably won't produce anything better?
| Tossrock wrote:
| It's true - the Lyapunov exponent shows that even arbitrarily
| close points in the system's phase space become separated by
| exponentially larger distances in time. So even with a
| computer the size of the universe, you can't really go
| further than 14 days. I'd highly recommend this Omega Tau
| podcast episode if you're interested in hearing more about
| chaos and predictability:
|
| http://omegataupodcast.net/119-chaos/
| spdustin wrote:
| That's an interesting paper (if a little lightweight) but it's
| begged the question: why is temperature measurement variance
| the rubric for evaluating weather models? (from footnote on
| page 9): "Consistent with the norms in this field, only the
| error in the prediction of maximum and minimum temperature is
| shown, but this result holds when we use other temperature
| indicators such as average temperature."
|
| Any meteorologists on HN able to weigh in?
| danielmarkbruce wrote:
| I mean...it's surprising this is a paper isn't it? Atoms, mass,
| charges, energy, forces etc are largely understood. The problem
| with weather, protein folding, oil exploration (and many other
| things simulations are run to predict) has always been that you
| can't do enough calculations in any reasonable amount of time
| (/money) so you have to figure short cuts which are
| approximately right. It's the same as graphics.
|
| It's self evident that the answer to a lot of these things is
| just "more compute" and "better shortcuts". Like, GPUs and deep
| neural nets.
| fallat wrote:
| Where 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 wrote:
| The 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.
| jsight wrote:
| Yeah, it is as if some group forgot the new "Be Evil" motto.
| ctoth wrote:
| My 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 wrote:
| It 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
| loxias wrote:
| I 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.
| kposehn wrote:
| I 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 wrote:
| It's already happening. For example, ECMWF provides
| experimental forecasts by their own deep learning model, plus
| models from Deepmind, NVIDIA, and Huawai.
| ckrapu wrote:
| Cool 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 wrote:
| All this research from Google and _still_ no discussion of using
| their access to billions of barometers from Android devices?
| _sigh_.
| spdustin wrote:
| Here's the paper about MetNet-3:
|
| https://arxiv.org/abs/2306.06079
| the_doctah wrote:
| When I search with my Google app it says "Source: weather.com"
| YossarianFrPrez wrote:
| How 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.
___________________________________________________________________
(page generated 2023-11-03 23:00 UTC)