Slashdot is powered by your submissions, so send in your scoop

 



Forgot your password?
typodupeerror
×
AI

How AI Can Make Weather Forecasting Better and Cheaper (bloomberg.com) 23

An anonymous reader quotes a report from Bloomberg: In early February a black box crammed with computer processors took a flight from California to Uganda. The squat, 4-foot-high box resembled a giant stereo amp. Once settled into place in Kampala, its job was to predict the weather better than anything the nation had used before (Warning: source may be paywalled; alternative source). The California startup that shipped the device, Atmo AI, plans by this summer to swap it out for a grander invention: a sleek, metallic supercomputer standing 8 feet tall and packing in 20 times more power. "It's meant to be the iPhone of global meteorology," says Alexander Levy, Atmo's co-founder and chief executive officer. That's a nod to Apple's design cred and market strategy: In many countries, consumers who'd never owned desktop computers bought smartphones in droves. Similarly, Atmo says, countries without the pricey supercomputers and data centers needed to make state-of-the-art weather forecasts -- effectively, every nation that's not a global superpower -- will pay for its cheaper device instead.

For its first customer, though, the Uganda National Meteorological Authority (UNMA), Atmo is sending its beta version, the plain black box. Prizing function over form seems wise for the urgent problem at hand. In recent years, Uganda has had landslides, floods, and a Biblical plague of locusts that devastated farms. The locusts came after sporadic drought and rain, stunning officials who didn't anticipate the swarms. "It became an eye-opener for us," says David Elweru, UNMA's acting executive director. Many nations facing such ravages lack the most modern tools to plan for the changing climate. Atmo says artificial intelligence programs are the answer. "Response begins with predictions," Levy says. "If we expect countries to react to events only after they've happened, we're dooming people to disaster and suffering." It's a novel approach. Meteorology poses considerable challenges for AI systems, and only a few weather authorities have experimented with it. Most countries haven't had the resources to try.

Ugandan officials signed a multi-year deal with Atmo but declined to share the terms. The UNMA picked the startup partly because its device was "way, way cheaper" than alternatives, according to Stephen Kaboyo, an investor advising Atmo in Uganda. Kaboyo spoke by phone in February, Kampala's dry season, as rain pelted the city. "We haven't seen this before," he said of the weather. "Who knows what is going to happen in the next three seasons?" [...] Atmo reports that its early tests have doubled the accuracy scores of baseline forecasts in Southeast Asia, where the startup is pursuing contracts. Initial tests on the ground in Uganda correctly predicted rainfall when other systems didn't, according to UNMA officials.

This discussion has been archived. No new comments can be posted.

How AI Can Make Weather Forecasting Better and Cheaper

Comments Filter:
  • Whatever they are doing, they are not serious about it's long-term viability.

    Atmo borrowed a liquid cooling technique developed for Bitcoin mining for its ventilation system.
    ...
    Atmo says it hopes customers will display it outdoors like a sculpture.

    These are two things that promote hardware failures.

    • by Entrope ( 68843 )

      They want to be the iPhone of weather forecasting: an overpriced style statement that locks you into one ecosystem.

      What I don't get is why their first customer is Uganda.

  • All it has to do is say there is a 50/50 chance of rain and it'll be doing far better than the crap forecasts that the met office does.

    • I would like to use this occasion to promote my new AI enhanced communicationsystem. It has almost zero latency, range is unlimited, it could be used to communicate with mars easily. Throughput of the first generation is 10GBps bidirectional*. No new hardware required. A simple wifi antenna suffices.
      *The catch? The bit-errorrate is 50%. So you will need to send everything twice.
    • All it has to do is say there is a 50/50 chance of rain and it'll be doing far better than the crap forecasts that the met office does.

      True story. My college room mate was a Meteorological major in college at UCLA. One of the classes he took was (unsurprisingly) "Weather Forecasting" where the whole focus of the class was to make a weather prediction for the Los Angeles area during the class. The class curve was based on your prediction accuracy. As a baseline, they had a "dummy" student that had as its prediction that the weather would be "Sunny with a high of 78F".

      At the end of the class, the "dummy" student's baseline weather predicti

  • "It's meant to be the iPhone of global meteorology,"

    It breaks very easily?

  • We run the danger of relying on AI to give us answers without knowing why it is giving us those answers, that kind of strikes me as having negative unknown consequences. One consequence includes there being less study of other forecasting methods or recognition methods if AI can do a better job.

    • Physics-based models feel like they are on more solid ground, but are they? We know for a fact we can't collect the data to fully specify all the starting conditions, and that we haven't identified all the phenomena that could affect the forecast, and that we can't make one integrated model of all the factors we do know about. So like any theory-based enterprise, it's nice and rigorous - right after you make a bunch of simplifying assumptions that may or may not totally invalidate the results.
      • by Kiliani ( 816330 )

        Hm ... if a physical model contains all the relevant relationships, then we don't need to "identified all the phenomena that could affect the forecast" - the phenomena should *reveal themselves* if the model has all the relevant physics in it. In reality, there are, if course, many issues: incomplete data, models leaving relevant physics out (e.g., because it was deemed not relevant enough, because it is computationally prohibitive, or because it does not potentially show effects in the grid sizes chosen),

    • by dvice ( 6309704 )

      > We run the danger of relying on AI to give us answers without knowing why

      Why would we not know why? You can train AI to give results where it explains why it came into the conclusion. This has been done several times. We have even been able to identify new 4th method for searching cancer from images based on the output from the AI.

  • More weather stations are needed. None of the weather services seem to have any station in the small town I live in; it seems as though they're interpolating from the towns north and south of me. While better than nothing, this does not produce accurate results. This is poignant today, where we are on the cusp of the rain/snow line in this storm.

  • AI would be great!
    I'm tired of all those Weather Critters who constantly lie to us about the weather. Who needs them?
    I already have family, friends, co-workers, and government personnel to lie to me. Why do I want the TV to lie to me?
    • by dvice ( 6309704 )

      "Do I look fat?" If you ask me, I will tell her the truth. And I can tell you that people actually want others to lie to them. Also when people tell you about their problems you would assume that they do it to hear some solutions, but in reality they don't. They just want to hear some bs words that make them feel better.

      For these same reasons, if you are a normal person, you want the TV to lie to you. You just can't handle the truth.

      • by tsqr ( 808554 )

        when people tell you about their problems you would assume that they do it to hear some solutions

        You'd be wrong most of the time. Generally, unless they specifically ask for your advice, people who tell you their problems are just seeking a sympathetic ear./p

  • The problem with weather forecasting is that it requires immense amounts of real time data to make accurate predictions, and immense amounts of compute power to crunch the data. The reason that large amounts of data are needed is that the system is very sensitive to small variations.

    If by "AI" is meant deep learning, that is a statistical technique that relies on training data to accurately reflect reality, and for the parts of the space of possibilities that are not seen in training to be well-behaved. T

    • by q_e_t ( 5104099 )
      80 years ago, weather prediction was basically done via statistical tables. It could produce reasonable results, but could also be thrown by them. This is presumably effectively a return to statistical tables, but with much more detail and somewhat better results. It then becomes a question of whether 90% accurate results for $10,000 a prediction is fine compared to, say, 95% accurate for $1,000,000.
  • emitting a loud hum from its whirring processors

    It seems to use those rejected, whirring second-rate processors instead of the nice humming first rate ones.

"Your stupidity, Allen, is simply not up to par." -- Dave Mack (mack@inco.UUCP) "Yours is." -- Allen Gwinn (allen@sulaco.sigma.com), in alt.flame

Working...