search

m2lines / L96_demo

Lorenz 1996 two time-scale model for learning machine learning
https://m2lines.github.io/L96_demo
Other
38 stars 20 forks source link

Team meetings topics suggestions #10

Closed johannag126 closed 3 years ago

johannag126 commented 3 years ago

Please suggest topics for team meetings below. If you see an already suggested topic that interests you, give it a thumbs up.

johannag126 commented 3 years ago

Learning about different NN flavors (ResNET and RNN, etc ) and their sensitivity to develop an intuition with a common baseline (L96 or otherwise)

johannag126 commented 3 years ago

Deep learning vs. other ML methods (symbolic regression, SinDY for equation discovery)

johannag126 commented 3 years ago

Metrics: learning subgrid (e.g., MSE of forcing or something else) + assessing climate models (e.g., climatology of SSTs, PDFs, ...)

judithberner commented 3 years ago

I would be interested in short presentations where an unrepresented process has been Mlearned. Laure's work is an example. Have there been successes in algorithmic learning e.g. in the PBL or other physical parameterizations? This is related to "Metrics: learning subgrid + assessing climate models" bit goes beyond the metric.

adcroft commented 3 years ago

Moving beyond L96 to a model with spatial dimensions (e.g. SWE or QG)

judithberner commented 3 years ago

Moving beyond L96 to a model with spatial dimensions (e.g. SWE or QG) Big thumbs up. We might want to consider if we want a -3 or -5/3(ish) spectrum, so maybe SQG?

judithberner commented 3 years ago

I think Tarun brought up a good point. By definition estimating model-error from analysis increments focuses on "fast-physics" errors which are arguably the low hanging fruit. However, a big problem for coupled climate modeling is climate drift which might have to do with the coupling itself (coupling frequency, "flux-correction", in short - "not fast physics"). I think very little work has done on this problem, but it might not be the best one for this group to tackle.

dhruvbalwada commented 3 years ago

We might want to consider if we want a -3 or -5/3(ish) spectrum, so maybe SQG?

Could do them all : https://www.cambridge.org/core/journals/journal-of-fluid-mechanics/article/abs/relative-dispersion-in-generalized-twodimensional-turbulence/08FB0083BEF447F4E528555341687809 ;)

LaureZanna commented 3 years ago

Moving beyond L96 to a model with spatial dimensions (e.g. SWE or QG) Big thumbs up. We might want to consider if we want a -3 or -5/3(ish) spectrum, so maybe SQG?

At NYU, we are using pyqg to learn, test and develop ML parameterizations as part of our projects .

adcroft commented 3 years ago

Understood about active projects. What about Cane and Zebiak, 1987 as an intermediate model between L96 and an eddying problem?

LaureZanna commented 3 years ago

Understood about active projects. What about Cane and Zebiak, 1987 as an intermediate model between L96 and an eddying problem?

We can still use QG or something else. But maybe you want to explain which topic(s) you want to explore then we can pick a model based on that! Not all models are good for all questions...

adcroft commented 3 years ago

I'd like to see how equation discovery works, and how different spatial resolution affects the learned models. I wa assuming we'd need PDEs rather than ODEs. Maybe even a 1D model would work for this: say the transport problem...

From: Laure Zanna @.> Sent: Friday, August 20, 2021 12:45 PM To: m2lines/L96_demo @.> Cc: Alistair J. Adcroft @.>; Comment @.> Subject: Re: [m2lines/L96_demo] Team meetings topics suggestions (#10)

Understood about active projects. What about Cane and Zebiak, 1987 as an intermediate model between L96 and an eddying problem?

We can still use QG or something else. But maybe you want to explain which topic(s) you want to explore then we can pick a model based on that! Not all models are good for all questions...

— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/m2lines/L96_demo/issues/10#issuecomment-902821897, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ABMWR4ZZIFMYIPK5ACTUSBTT52BEFANCNFSM5CMTCZJA. Triage notifications on the go with GitHub Mobile for iOShttps://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Androidhttps://play.google.com/store/apps/details?id=com.github.android&utm_campaign=notification-email.

LaureZanna commented 3 years ago

When I started several years ago, I used a advection-diffusion for all equation-discovery problems (1D and 2D) to get an intuition. It was an eye opener to see the issues with different ML methodologies (SINdy in particular).

adcroft commented 3 years ago

I'm sold. Consider OP to read "advection-diffusion in 1d".

-- Dr Alistair Adcroft email: @.**@.> Program in Atmospheric & Oceanic Sciences Cal: linkhttps://calendar.google.com/calendar/embed?src=alistair.adcroft%40noaa.gov&ctz=America%2FNew_York Tel: (609) 987-5073 Princeton University, 300 Forrestal Rd, Sayre Hall, Princeton, NJ 08540-6654

From: Laure Zanna @.> Sent: Friday, August 20, 2021 1:35 PM To: m2lines/L96_demo @.> Cc: Alistair J. Adcroft @.>; Comment @.> Subject: Re: [m2lines/L96_demo] Team meetings topics suggestions (#10)

When I started several years ago, I used a advection-diffusion for all equation-discovery problems (1D and 2D) to get an intuition. It was an eye opener to see the issues with different ML methodologies (SINdy in particular).

— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/m2lines/L96_demo/issues/10#issuecomment-902849726, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ABMWR46CQFDUPNAIM6ADAULT52G7TANCNFSM5CMTCZJA. Triage notifications on the go with GitHub Mobile for iOShttps://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Androidhttps://play.google.com/store/apps/details?id=com.github.android&utm_campaign=notification-email.

judithberner commented 3 years ago

Oceanographers by heart !.... I would also like spatial scales and subgrid-scale processes with energy / enstrophy cascades.

LaureZanna commented 3 years ago

Oceanographers by heart !.... I would also like spatial scales and subgrid-scale processes with energy / enstrophy cascades.

yes we are. we can get our students and postdocs to present their work and maybe do an active hands on session on this topic for sure if they are up for it; this is active research ! I will post a few plots from @arthurBarthe in pyqg. This is more about update on research. We might combine these posts to clean up the topics and discussions. Have a good weekend!