Alright, so it has been a while since there has been updates on the single ESM calibration activities so for those who need a refresher here is goes…
We originally tried to find the Hector parameters that best fit temperature. Although we were able to solve for Hector parameters the values for S and diff returned were unrealistic, which we determined was caused by S and diff trading off with one another. When we were only using temperature data in the MSE S and diff were unidentifiable so we ended up adding heat flux as an additional comparison data.
When we use the heat flux data we ran into two issues.
Solution
So in order to avoid the obstacles that diff, S, and limited heat flux data present us we decided to use the range of the cmip range for the 2100 heat flux + the mesa function RPL designed for the MCMC as the tie breaker to use in the Hector calibration. We also do not fit volscl for the models that do not provide historical output because often times optim takes volscl to funky town even though it has little to no impact on the performance of Hector’s temperature fit in the future period.
So here are our results I think they look pretty good for the most part and have a few questions for you at the end.
For each model? What models do the best and the worst?
What is the model that Hector does the best job emulating?
## [1] "MRI-CGCM3"
What is the model that Hector does the worse job emulating?
## [1] "MPI-ESM-P"
What does the values we are trying to minimize look like for each variable / experiment? Is it consistent across models / variables / ensembles?
| variable | experiment | min | max | mean | sd | range |
|---|---|---|---|---|---|---|
| heatflux | rcp26 | 1.175 | 1.176 | 1.175 | 0.000 | 0.000 |
| heatflux | rcp45 | 1.088 | 1.088 | 1.088 | 0.000 | 0.000 |
| heatflux | rcp60 | 0.837 | 0.838 | 0.837 | 0.000 | 0.001 |
| heatflux | rcp85 | 1.286 | 1.287 | 1.286 | 0.000 | 0.001 |
| Tgav | historical | 0.037 | 1.802 | 0.365 | 0.260 | 1.765 |
| Tgav | rcp26 | 0.005 | 0.065 | 0.029 | 0.015 | 0.060 |
| Tgav | rcp45 | 0.003 | 0.114 | 0.033 | 0.028 | 0.112 |
| Tgav | rcp60 | 0.003 | 0.084 | 0.023 | 0.020 | 0.081 |
| Tgav | rcp85 | 0.005 | 0.075 | 0.025 | 0.019 | 0.071 |
As we expected (because of the way that the mesa function work) the min heatflux value is the same across the models/ensembles.
So it looks like there are still some unreasonable values of diff, alpha, and volscl.
## # A tibble: 4 x 4
## param min max mean
## <chr> <dbl> <dbl> <dbl>
## 1 alpha 0.119 2.45 1.48
## 2 diff 1.38 788. 42.8
## 3 S 1.23 5.38 3.56
## 4 volscl -0.373 3.95 1.18
Which have unreasonable param values?
For alpha? (This is going to be values out side of the range of 0 to 1)
| model | S | diff | alpha | volscl |
|---|---|---|---|---|
| ACCESS1-0 | 4.481406 | 4.022459 | 2.169738 | 1.3063726 |
| ACCESS1-3 | 4.137873 | 4.869920 | 2.012696 | 1.3027606 |
| CanESM2 | 5.379942 | 2.676373 | 1.845813 | 1.4425507 |
| CCSM4 | 3.417640 | 2.150772 | 1.361184 | 0.8032603 |
| CESM1-BGC | 4.212259 | 4.757437 | 1.254342 | 1.1186954 |
| CESM1-CAM5 | 4.672884 | 3.511973 | 2.004213 | 1.0839185 |
| CMCC-CESM | 3.066088 | 2.427470 | 2.454835 | 0.1542517 |
| CMCC-CM | 3.344814 | 2.104657 | 1.797835 | -0.3734199 |
| CMCC-CMS | 3.492108 | 1.933353 | 1.819106 | -0.1699221 |
| CNRM-CM5 | 3.793208 | 6.027884 | 1.340480 | 1.1009000 |
| CSIRO-Mk3-6-0 | 4.160642 | 5.036721 | 1.662681 | 1.6443379 |
| CSIRO-Mk3L-1-2 | 2.832724 | 11.630328 | 1.152573 | 0.9342279 |
| EC-EARTH | 4.288184 | 4.615576 | 1.925060 | 1.4865312 |
| FGOALS-g2 | 2.603372 | 3.416440 | 1.866707 | -0.0187293 |
| GFDL-CM3 | 3.925983 | 1.375715 | 2.226984 | 0.9854006 |
| GFDL-ESM2G | 2.205985 | 5.230977 | 1.360018 | 1.7792677 |
| GISS-E2-R-CC | 3.033876 | 10.034147 | 1.068817 | 1.3610505 |
| GISS-E2-R | 3.273466 | 7.933532 | 1.226466 | 1.4798282 |
| inmcm4 | 2.816547 | 11.726621 | 1.160714 | 0.6170523 |
| IPSL-CM5A-MR | 4.760392 | 3.727721 | 1.142407 | 1.1579202 |
| MIROC-ESM-CHEM | 5.377651 | 2.763389 | 1.768058 | 0.9420724 |
| MIROC-ESM | 5.283901 | 2.968408 | 1.613244 | 1.3916866 |
| MIROC5 | 3.640256 | 6.619373 | 1.484628 | 1.5105629 |
| MPI-ESM-LR | 3.231097 | 2.406721 | 1.499654 | 0.8941075 |
| MPI-ESM-MR | 3.184624 | 2.382020 | 1.812322 | 0.9785344 |
| MRI-CGCM3 | 3.207000 | 8.602656 | 1.733598 | 1.2702199 |
| MRI-ESM1 | 1.228837 | 788.170015 | 1.606054 | 2.9912901 |
| NorESM1-ME | 2.847869 | 2.789965 | 1.910678 | 0.3259734 |
Alright so it looks like the majority of the alpha param values are outside this range. Is that alarming?
For S? (we will say that the normal range is between 0 and 10)
## [1] model S diff alpha volscl
## <0 rows> (or 0-length row.names)
Well it looks like all of the S values are reasonable!
For diff? (we will say that the normal range is between 0 and 8)
## model S diff alpha volscl
## 1 CSIRO-Mk3L-1-2 2.832724 11.630328 1.1525729 0.9342279
## 2 GISS-E2-R-CC 3.033876 10.034147 1.0688171 1.3610505
## 3 inmcm4 2.816547 11.726621 1.1607136 0.6170523
## 4 MPI-ESM-P 1.464755 473.694144 0.1193029 3.9530229
## 5 MRI-CGCM3 3.207000 8.602656 1.7335980 1.2702199
## 6 MRI-ESM1 1.228837 788.170015 1.6060545 2.9912901
Okay so I think that the most alarming diff values are MRI-ESM1 and MPI-ESM-P. But I suspect that this is because these models only reported historical temperature data. May be we cannot fit to only historical temp + 2100 heat flux.
I am not surprised that inmcm4 is in there because it has wonky heat flux data. The other models with diff fits that range from 8 to 11 would need a bit more digging into.
Calibrated Hector pretty consistently underestimates the ESM temperature data around 1960.
This fit in not great, the temp over the historical period ramps up faster than I would have expected compared to the other ESM historical data. So it is not unsurprising to me that Hector struggles to be as warm as the ESM data.
There is almost no volcano dips in this historical period.
I think that the RCP 2.6 ESM output looks funky. SO I am not surprised that Hector over estimates RCP 2.6 temp from 2050 to 2100.
This is the model that Hector does the worst job of emulating. Which is not surprising since there is only the historical data to use as comparison data.
So this is the model with the smallest MSE value.
This diff value is one unrealistic but that is not surprising since the only ESM temp data we have is from the historical period.
It looks like most of the Hector values from the best fits fall at the ends of the cmip range.