Margaret Duffy (NCAR): "Perturbing parameters to understand cloud contributions to climate change"
Paulo Ceppi (Imperial College London): "Observational constraints on low cloud feedback"
Andrew Williams (Univ. of Oxford): "Circus tents, convective thresholds and the non-linear climate response to tropical SSTs"
00:28:22 Ceppi, Paulo A: Can we find observational evidence for this?
00:29:05 Tianle: how long was the perturbation experiment run, esp. the cooling ones?
00:29:27 Gavin Schmidt: What do we get if we apply the greens function approach to the 2000+ SST? Do we recover the CERES TOA changes?
00:30:35 ross herbert: Nice talk Andrew! Is this going to be sensitive to the base convective state of the region? Might the inflexion point will be dependent on this?
00:31:39 McKim, Brett: Why is the convective threshold defined with respect to the local h_ft and not h_ft from regions of deep convection? Papers by Yi Zhang have shown that h_ft in regions of deep convection are remarkably uniform.
00:32:05 Cristian Proistosescu: @gavin: figure 13 in Zhou et al (at least for CRE). https://doi.org/10.1002/2017MS001096
00:32:32 Gavin Schmidt: Reacted to "@gavin: figure 13 in..." with 👍
00:32:39 Jonah Bloch-Johnson: @Gavin - if we apply the Green’s function method to SST patterns from the last few decades, we can recreate the atmospheric model’s response; but for many models it struggles to get the response to the abrupt4x (see also https://www.authorea.com/users/554553/articles/627762-the-green-s-function-model-intercomparison-project-gfmip-protocol)
00:32:47 Brian Rose: Does the metaphor extend to an explanation for the “tautness” of the tent fabric? Is this related in a simple way to the deformation radius of the tropical atmosphere?
00:32:49 Jonah Bloch-Johnson: *to the abrupt4x pattern
00:34:04 Andrew Williams (he/him): Hi Tianle! All experiments run for 10 years, except for the control which was run for 30 years.
00:34:05 Cristian Proistosescu: Reacted to "Does the metaphor ex..." with 💯
00:35:44 Andrew Williams (he/him): Replying to "Nice talk Andrew! Is..."
Thanks Ross! Yep, definitely. If a region is only weakly convecting then a small, negative delta_SST would be enough to shut it down. This makes the Central Pacific a really sensitive region to SSTs generally
00:36:56 Cristian Proistosescu: Replying to "Nice talk Andrew! Is..."
This suggests ENSO might be a good sandbox to test this theory?
00:37:11 Andrew Williams (he/him): Reacted to "This suggests ENSO m..." with 👍
00:37:21 Andrew Williams (he/him): Replying to "Nice talk Andrew! Is..."
Yep! good point!
00:37:33 Cristian Proistosescu: Replying to "Nice talk Andrew! Is..."
Related, would this also explain why convection shifts during ENSO? - you’ve raised a higher tentpole in the East Pacific?
00:37:44 Andrew Williams (he/him): Reacted to "Related, would this ..." with 👍
00:40:29 Andrew Williams (he/him): Replying to "Why is the convectiv..."
we touch on this a bit in the paper, but in short:
Our results are not massively different if we use the tropically averaged h_ft (or, roughly equivalently, the h_ft in regions of deep convection). However, the tropical-avg h_ft is only equal to the h_ft in regions of deep convection is WTG holds perfectly. In reality you do get zonal temperature gradients in the tropical free-troposphere (cf Bao et al., 2022 JClim), which are ~O(3K). Understanding the pattern effect also involves understanding the response to perturbations of ~O(3K), so it becomes important (at least theoretically) to retain the local h_ft
00:40:39 Andrew Williams (he/him): Replying to "Why is the convectiv..."
Happy to chat offline, it’s a bit of a subtle point i think
00:41:12 Gavin Schmidt: How does this tie in to the CESM2 code fixes highlighted by Zhu et al 10.1029/2021MS002776
00:42:06 Andrew Williams (he/him): Replying to "Does the metaphor ex..."
Hi Brian! Essentially, yes. We relate the `tautness` of the tent fabric to the strength of the WTG constraint in the tropics. Because gravity waves are very effective at spreading out buoyancy gradients, you can’t get very strong gradients in h^{*}_ft, which is equivalent to a ‘’tight/taut’’ fabric.
00:42:27 Gavin Schmidt: Replying to "How does this tie in..."
Maybe that is already in v6.3?
00:42:47 ross herbert: Replying to "Does the metaphor ex..."
And the stripes are gravity waves?
00:43:36 Andrew Williams (he/him): Replying to "Can we find observat..."
Hi Paulo, hope my answer made sense for this? There is observational evidence for the threshold behaviour, and so the other components that would need observational constraints on would be WTG and the link between inversions/low clouds/TOA
00:43:51 Andrew Williams (he/him): Replying to "Does the metaphor ex..."
Yep! Ooops, I may have forgotten to say that...
00:43:59 Ceppi, Paulo A: Reacted to "Hi Paulo, hope my an..." with 👍
00:44:19 Andrew Williams (he/him): Replying to "Does the metaphor ex..."
Oh wait, you’re trolling Ross? ;)
00:44:26 Andrew Williams (he/him): Reacted to "And the stripes are ..." with 🎪
00:46:36 Mark Zelinka: @Margaret: nice talk! It is interesting to me that CAM6 has a fairly strong negative high latitude low cloud optical depth feedback. I thought Bjordal et al showed that this model’s extratropical optical depth feedback was positive (or at least eventually becomes positive)
00:47:15 McKim, Brett: Replying to "Why is the convectiv..."
Thanks!
00:47:36 Gavin Schmidt: How did you do the sampling in the 45-dimension parameter space?
00:48:09 Karsten Haustein: @Margaret: is high-cloud altitude referring to the altitude cirrus level? guess i’m being lost in translation ;) … and is the relatively high uncertainty related to the strength of the convective response to warming?
00:48:58 Marcus van Lier-Walqui (he/him): Is the sensitivity analysis linear? do you consider nonlinear or pairwise interactions? Cool work!
00:48:59 Thorsten Mauritsen: Great work Margaret! I have a philosophical question. Now that you have a model on how you can tune ECS in the model, has NCAR given thought to how you will use this information in the future?
00:49:04 ross herbert: Replying to "Does the metaphor ex..."
Haha I was! Nice
00:49:17 dennishartmann: Interesting work. Could you explain again why cloud parameters produce the spread and yet are not responsible for the change in sensitivity from CAM4 to 6.
00:49:42 Brian Rose: Replying to "Does the metaphor ex..."
Great, thanks!
00:51:20 Gavin Schmidt: Thanks.
00:51:25 Virendra Ghate: Reacted to "Interesting work. Co..." with 👍🏽
00:51:59 Kevin Smalley: Reacted to "Interesting work. Co..." with 👍🏽
00:52:36 Cristian Proistosescu: Reacted to "Great work Margaret!..." with 💯
00:52:40 Gavin Schmidt: But you can use this to tune to real observables, not ECS, surely?
00:52:52 Duncan Watson-Parris: Reacted to "But you can use this..." with ➕
00:53:00 Andrew Williams (he/him): Reacted to "But you can use this..." with 👍
00:53:00 dennishartmann: Don’t folks already tune parameters to get what they want?
00:53:10 Cristian Proistosescu: Reacted to "Don’t folks already ..." with 🤫
00:53:21 Gavin Schmidt: Replying to "Don’t folks already ..."
Not ECS (at least in general).
00:53:49 yokotsushima: Do you have any idea which structural change contribute to a weaker feedback?
00:53:57 Thorsten Mauritsen: Thanks Margaret. It is hard to ignore that you have this option. In the past at MPI we decided to tune the historical warming with similar information at hand.
00:55:24 Duncan Watson-Parris: Thanks for the really nice talk Margaret. I particularly like the assessment of the parameter differences between CAM5 and CAM6. How robust is the linearity assumption you have to make though? Could you use another regression approach (such as a Gaussian Process)?
00:56:32 Margaret Duffy: Replying to "@Margaret: is high-c..."
Thank you for your question! This is the same high cloud altitude feedback as in Sherwood et al. and calculated in Zelinka et al. 2022. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2021JD035198
00:56:52 Gavin Schmidt: Tuning to uncertain things (ECS or historical warming | uncertain forcing) risks overfitting. Sampling from a reasonable distribution though… that’s reasonable.
00:57:11 dennishartmann: Replying to "Don’t folks already ..."
I think if a model produced a really wild sensitivity, and could not match the observations, an easy fix would be to look at the cloud parameters, since they are so uncertain anyway.
00:58:21 Gavin Schmidt: Replying to "Don’t folks already ..."
And yet… https://www.realclimate.org/images/cmip_ecs_Aug9-1536x1083.png
00:58:26 Margaret Duffy: Replying to "Interesting work. Co..."
Yes, my work both highlights the sensitivity to parameters, yet says that parameters are not responsible for the increase in cloud feedback/ECS from CAM5 to CAM6. The short answer is that I think both structural and parametric sensitivity can affect the cloud feedback. Longer answer is that there are quite a few structural changes to the model from CAM5 to CAM6, including the new CLUBB turbulence scheme, which means that many of the parameters in CAM6 aren’t even in CAM5!
00:58:37 Jonah Shaw (he/him/his): Hi Margaret, great talk! You narrowed the simulations down using the cloud radiative effect. I am wondering how well these 206 simulations capture the observed mean state of cloud amount and cloud phase? Do your results change if you only evaluate simulations that are reasonably capturing the observed cloud mean state (in addition to the CRE)?
00:59:03 Duncan Watson-Parris: Reacted to "Yes, my work both hi..." with 👍
00:59:36 Eric DeWeaver: Replying to "Thanks Margaret. It ..."
Are there opportunities to look at the consequences of the tuning values for real-world effects that could be compared against observations? For instance do they have consequences for cloud behavior during ENSO events that could be compared against satellite obs?
00:59:51 Karsten Haustein: Replying to "@Margaret: is high-c..."
👍
01:00:14 Margaret Duffy: Replying to "But you can use this..."
Yes this is an interesting question. When I was looking at the difference in clouds between CAM6.0 and CAM6.3 I compared to obs and the difference between the model and observations were so much larger than the difference between the two generations of the model, which I thought discouraging.
01:01:42 Margaret Duffy: Replying to "Do you have any idea..."
Thank you for your question. I do not know for sure. My colleagues and I think it is probably changes to either CLUBB (turbulence) or MG (microphysics), but can’t say anything more specific at this time.
01:01:51 Gavin Schmidt: Replying to "But you can use this..."
We’ve been using something very similar for the tuning of our recent GISS-E3 model. It works well - and produces multiple ‘equally good’ parameter sets that have different ECS. Talk to Greg Elsaesser perhaps if you are interested.
01:02:41 Thorsten Mauritsen: Replying to "Thanks Margaret. It ..."
@eric, yes you could. It is hard to serve multiple purposes, though. Number one tuning target is radiation balance to get a stable climate, and after that you typically go down a list of priorities. At MPI we found one parameter useful for tuning ENSO amplitude, but of course its right for the wrong reasons.
01:03:01 Margaret Duffy: Replying to "Thanks for the reall..."
Hi Duncan. Yes! Your emulator work has been helpful here. I used a regularized regression and found similar MSE using regression as compared to using emulators, so I used the regression for simplicity and interpretability
01:03:15 dennishartmann: Replying to "Don’t folks already ..."
And that is the result AFTER tuning, so plenty of work to do yet.
01:04:09 Marcus van Lier-Walqui (he/him): Replying to "But you can use this..."
+1 an important “secret sauce” to the GISS tuning process has been dealing with observational uncertainties, including structural errors that characterize the model’s inability to match observations for *any* parameter combinations (Greg’s put time into this)
01:05:02 Margaret Duffy: Replying to "Don’t folks already ..."
Hi, I think some modeling centers may tune to ECS. NCAR does not. You are right that our ECS is very sensitive, though.
01:05:30 Duncan Watson-Parris: Replying to "Thanks for the reall..."
Nice! Yes, if linear models work they’re definitely preferable.
01:05:34 Duncan Watson-Parris: Reacted to "Hi Duncan. Yes! Your..." with 👍
01:06:45 ross herbert: Really interesting Paulo, do you have any idea whether the optical depth differences are due to LWP or effective radius? Or both!
01:06:46 Andrew Williams (he/him): Very cool stuff! If cloud feedbacks are stronger over the low cloud/subsidence regions than models predict, does that mean we should be less trusting of the SST warming patterns predicted by coupled models under abrupt 4xCO2? (i.e. fairly uniform/El-Nino like warming)?
01:06:49 Joel Norris: It's kind of tricky to distinguish cloud fraction changes from optical thickness changes in observations because it depends on whether optically thin/partial cloud covered pixels are counted or not
01:06:49 Margaret Duffy: Replying to "Hi Margaret, great t..."
Hi Jonah. Yes, I glazed over this but I had color shading on the plot for a base-state cloud error metric from a Stephen Klein paper. Our TOA imbalance metric eliminated the unskilled simulations based on base-state cloud errors, too. Reassuring.
01:07:07 Gavin Schmidt: Replying to "Don’t folks already ..."
This might make a good discussion for another seminar in this series!
01:07:32 Virendra Ghate: Reacted to "It's kind of tricky ..." with 👍🏽
01:08:08 Tianle: Reacted to "It's kind of tricky ..." with 👍
01:08:32 Jesse Loveridge: Reacted to "It's kind of tricky ..." with 👍
01:08:51 Margaret Duffy: Replying to "But you can use this..."
Yes. This is a complicated subject but I’ll just add that I tended to find moderate cloud feedbacks for parameter combinations which gave good results as compared to Sherwood/WCRP
01:09:08 Isaac Held: Paolo, it is impressive how the models “historical” sensitivities reproduce the force response. Are these sensitivities just from high frequencies? do they include ElNino?
01:09:31 dennishartmann: @Paulo Great work. Joel Norris asked my question. Cloud fraction is a concept, dependent on a threshold.
01:09:57 Eric DeWeaver: Replying to "Thanks Margaret. It ..."
Thanks. I suppose the question is whether there are any emergent constraints that could rule out certain parameter settings.
01:09:58 Rodrigo Caballero: Nice talk Paulo. What time averaging do you use for the CCF analysis? Is there sensitivity to that (ie monthly vs daily)?
01:11:29 Yuan-Jen Lin: Is the overall (base-state) optical depth different in observation different from models' historical *simulations*? And if that would contribute to dR/dX biases?
01:11:42 Mark Webb: Hi Paulo, nice work. Some of the CGILS results did show low cloud thinning which GCMs may fail to capture due to their causer vertical resolution:https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1002/2016MS000765
01:11:59 Tianle: Paulo, great work. What was the rational behind the assumption that high ecs models are overestimating?
01:12:12 Margaret Duffy: Replying to "Thanks Margaret. It ..."
Hi all. This is a big question and I won’t say anything definitive. I’ll just say that I did a constraint of differences in the PPE as compared to WCRP/Sherwood values and found moderate cloud feedbacks in low error models. There is a set of parameters that robustly gets this. BUT I am not necessarily advocating for tuning to WCRP/Sherwood.
01:12:44 Yuan-Jen Lin: thank you @Paulo!
01:12:47 Ceppi, Paulo A: Reacted to "Hi Paulo, nice work...." with 👍
01:12:49 Larry Di Girolamo: @Dennis, the reported optical depth by MODIS is dependent on the cloud detection thresholds AND other quality control factors that reduce the representation of the data.
01:13:04 Ceppi, Paulo A: Reacted to "@Dennis, the reporte..." with 👍
01:13:09 Tianle: Reacted to "@Dennis, the reporte..." with 👍
01:15:10 Gavin Schmidt: We can tell.
01:15:31 Tianle: Reacted to "We can tell." with 😃
01:15:40 Gavin Schmidt: Thanks!