"without using a supercomputer", you say? Then what was used to make that precious training and validation dataset, ERA5? This is not a criticism of AI models, but the way their advances are currently being reported www.nature.com/articles/d41... - ThreadSky

simonleewx.com • 129 days ago

"without using a supercomputer", you say? Then what was used to make that precious training and validation dataset, ERA5?

This is not a criticism of AI models, but the way their advances are currently being reported https://www.nature.com/articles/d41586-024-02558-4

Comments

70dbz.bsky.social•128 days ago

Precip is still not great, yet.

pppapin.bsky.social•128 days ago

I’m not convinced it will ever be “great” unless it’s trained off a completely different dataset (WSR-88d archive?).

vortexjeff.bsky.social•128 days ago

Winter precipitation types may be even worse for AI. Observations do not differentiate between sleet and snow accumulations and ASOS frequently misrepresents sleet as light snow. How can you train AI on that?

clarkevanswx.bsky.social•128 days ago

The general lack of training on precipitation fields for the current-generation ML models represents an opportunity for improvement (by starting to do it!) and also an indictment of their current fidelity for representing physical processes, imo.

djgagnedos.bsky.social•128 days ago

Already happening. AIFS and Pangu 2 have precip outputs. ACE2 from Allen AI has more explicit physics conservation and more micro physical variables. https://arxiv.org/abs/2411.11268

clarkevanswx.bsky.social•128 days ago

Thanks for sharing. I’ve naively wondered if these conservation aspects, KE spectra, etc. could be incorporated into loss functions to improve the models’ physical consistency, but the module approach of the ACE2 team seems promising (and would better enforce these constraints).

jwheare.com•129 days ago

I’d love to see all the hallucinated attempts before they finally got something accurate.

stevie-zero.bsky.social•129 days ago

You wouldn’t need a supercomputer to make training and validation datasets, surely?

But one might be required to generate the model(s) and run them.

marc-alessi.bsky.social•129 days ago

There is a lot of merit in not running a supercomputer every time you need to make a forecast though… agreed it’s overdone by media, but only a bit

simonleewx.com•129 days ago

That's true, but reanalyses aren't a trivial effort yet seem to be reported as such – and ERA5 is often spoken about as though it's perfect (it isn't)

marc-alessi.bsky.social•129 days ago

Completely agree, and I don’t think ML forecasts would ever replace dynamical… dynamical models had huge wins this year being able to predict extreme weather events influenced by insane boundary conditions caused by climate change. Can’t see ML models beating that

heatwavekgns.bsky.social•128 days ago

AI/ML clearly early in its development. This abstract states that AI climate models trained on current climate demonstrates skill in warm climate scenarios as well as preindustrial, yet when run in a warmer or preindustrial climate, some models trend back to our current climate…

wbmol.bsky.social•126 days ago

Typical Nature article, too, like the original paper. Don't get me wrong, it's impressive and exciting work, but the language is overblown and limitations are marginalized... (Eg, precip, or like you say, ERA5 taken for granted)

daveosw16.bsky.social•129 days ago

It’s quite an interesting area of prediction. I’m aware some AI systems seem better at diagnosis than physicians in some cases. There, they use imagery which makes me wonder why couldn’t satellite images as well as temp, wind etc be used in AI models?

cjmoose.bsky.social•129 days ago

Seems to me like there are far fewer variables in a medical diagnosis of a single individual than predicting the chaotic movement of the particles in a massive three-dimensional fluid on a rotating body with significant daily flux of energy input.

simonleewx.com•129 days ago

Yes, that

daveosw16.bsky.social•129 days ago

Looking at mammographs was one area where AI did better than radiologist assessments, I believe. Shadowy areas, not easy to spot by eye, but seemingly AI could predict malignancy better. This is more a perceptual skill than finding symptoms to determine a diagnosis. There is some chaos in there.

weatherbadger.bsky.social•129 days ago

Headline written by ChatGPT?

Comments

Posting Rules

Reply