To-Do for Model

[kevinkuruc]

@FrankErrickson , I've done some work applying the model, but I think before we move to finalizing figures (and therefore a paper draft) there are a few things we need to do.

1. Make sure we've converted DICE to an annual model correctly. I think I noticed that your TFP interpolation starts with 2010 on the spreadsheet despite our model starting at 2015---if I read that right then every year will be quite a bit poorer than in DICE.

2. Find other sources that make us confident our FAIR results are reasonable (the biggest being that BAU peak temperature (5.5) is about 1.5 degrees lower than in DICE (7)). If you could also use what you know about Methane to make sure our pulse look sensible.

3. One reason I think something is off is that when I compute the SCC (shown in SCC_Notebook I've included in src_annualized folder) I get $26.5 in 2020 whereas the DICE2016 model Nordhaus reports hits $37. (See: https://www.pnas.org/content/pnas/114/7/1518.full.pdf). One possibility could just be that the FAIR model implies less warming per emission. That would explain both the temperature result and SCC result. We could also compute the social cost of methane to see if we get towards a number you know well. One possibility I'm looking into is that the GAMS version used for the 2017 PNAS paper is said to be different from the excel sheet (Nordhaus specifically notes the damage function on his website).

Let's collaborate to get these taken care of in the coming weeks, then Jon and I will get a draft polished ASAP!

[FrankErrickson]

@kevinkuruc A couple brief thoughts/questions. Happy to go back and forth to sort this out.

(1) Where do you see TFP starting at 2010? It looks like the Excel spreadsheet I read off of starts in 2015. In the annual_parameters.jl file, I have the line p[:tfp] = getparams(f, "B21:CW21", :all, "Base", T), which I thought corresponded to 2015. But very possible I indexed something wrong somewhere.

(2) So the FAIR model I coded up exactly matches the original Python version the core FAIR development team made (I have tests to show this). So we can be confident there aren't bugs in the climate model. Making a direct apples-to-apples comparison with FAIR vs. DICE will be a little tricky because they have a different structure, so even with the same emissions scenario you will have to make some assumptions about how comparable different parameterizations are. And also, we know the DICE climate model has some shortcomings so we should not expect them to give the same temperature response.

But one thing to note. The DICE2016 spreadsheet has a default equilibrium climate sensitivity of 3.1. The default setting for FAIR is 2.75. So not surprising they give different such different peak temps if this wasn't the same across models. One tricky issue with FAIR though is that it also has a separate parameter for the transient climate response (and this along with the ECS determine the short and long time scale responses of the climate system). DICE doesn't have this transient term, so again the apples-to-apples thing is complicated. But as a crude check, you could set DICE2016's ECS to 2.75 and see if they get a little closer.

(3) This could again be related to the climate sensitivity issue. With a lower climate sensitivity, each marginal pulse of gas produces less warming and less damages. And if the background temperature is also lower, then you are on a less steep part of the damage function. And at least for DICE2013 with the SC-CH4, I've never noticed huge differences between the DICE vs. GAMs versions unless you started getting into extreme temperature scenarios.

[FrankErrickson]

As a crude check, I just quickly ran baseline MimiDICE2016 with ECS=2.75 and peak temp went from 7.2 down to 6.45.

[kevinkuruc]

@FrankErrickson That all sounds good to me. I was only surprised about the SCC because I thought DICE was known for estimating a relatively low values of this, so for us to switch in a more sophisticated climate model and get something even smaller was surprising. DICE has revised up the SCC in recent iterations, so maybe that's not an accurate characterization of the field any longer. Alternatively, it could be that DICE has a low SCC for reasons unrelated to the climate module.

$27 still seems low to me, so Im going to make sure I've computed this term correctly by trying it in DICE2016 and seeing if I get back his PNAS results.

[kevinkuruc]

@FrankErrickson to continue our Skype conversation here, I've used my SCC calculation code on the DICE2016 model in its Mimi format. I get back the same $37 dollars that we're expectin; it seems we don't have a trivial SCC code bug. After our conversation I am more confident that the $26.50 I get using our model is just the result of FAIR (esp. with ECS of 2.75) implying less warming per emission, and hence less damages.

I still think it makes sense to go through and annualize DICE's climate model to make sure we get similar results in an annualized model. That will make us feel much better and it will make for a nice SI appendix. Something to keep in mind is that it may make sense to do a run where we set FAIR to give something equivalent to a 3.1 ECS parameter so that we can present results with this number in a main table. Since FAIR doesn't have an equivalent parameter, if you can (at some point) give me an idea of how to get into FAIR and change the right parameters, that'd be great.