The need to define terms like this is more evidence of climate science quickly developing a large service industry branch

Thanks to @pppapin.bsky.social @bmcnoldy.bsky.social and others for sharing their fascinating work on this

That ML systems are performing so well suggests they are learning these mappings v efficiently, but averaging also means they can miss extreme events, as noted in the article.

There's also the concern of whether ML systems will degrade as the climate moves outside the regime they were trained in

2. Physical models are run in ensembles to account for initial condition error. The goal is to reproduce a posterior PDF given uncertain ICs

ML models make predictions under the assumption that ICs are uncertain. I.e., they learn mappings from ICs to posterior PDFs that average over uncertainty...

1. Physics-based models have parameterizations of unresolved processes, which are a source of forecast error.

ML models implicitly have the same parameterizations, but these are tuned to whatever gives the most accurate forecast, and aren't constrained by underlying physical picture

So for climate purposes, hybrid models are reliable in many ways, but they can go off the rails quickly wherever they weren't explicitly trained.

Paper:
agupubs.onlinelibrary.wiley.com/doi/10.1029/...

NeuralGCM's response is comparable in many ways to conventional models (it might even be more reliable in some variables like blocking)...

But it performs poorly in the stratosphere, which wasn't a priority in tuning. This in turn affects strat-trop coupling as well

@robinsonmeyer.bsky.social makes the point that humans have always reshaped their environment, but the takeaway isn’t license to modify, it’s that communities have figured out ways to govern commons together (Ostrom, not Prometheus) (2/2)

...breaking the 2017 record for US weather damage (>$300 billion).

www.nytimes.com/2025/10/22/c...

Read the paper to learn much more about the mechanism, and watch out for follow-up work coming out soon.

Link: journals.ametsoc.org/view/journal...

To dig into these changes, Lily did a tour-de-force analysis, including:
-Cloud Controlling Factors
-Mixed-layer energy budgets
-Green's Function analysis

Together, these revealed how climate feedbacks respond to AMOC decline, with extratropical NA warming playing a key role

And his two recent Stirring Tropics papers with Victor Mayta, which give a new picture of the tropical circulation
doi.org/10.1175/JCLI...
doi.org/10.1175/JCLI...

Adames & Kim (2016) -- one of his first moisture mode/MJO papers
doi.org/10.1175/JAS-...

Adames & Wallace (2017) -- really careful analysis of the tropical atmospheric signature of El Nino
doi.org/10.1175/JAS-...