We covered this paper when it was first released, here is some commentary on it – Anthony
Guest essay by By PATRICK J. MICHAELS
On January 7 a paper by Veronika Eyring and 28 coauthors, titled “Taking Climate Model Evaluation to the Next Level” appeared in Nature Climate Change, Nature’s journal devoted exclusively to this one obviously under-researched subject.
For years, you dear readers have been subject to our railing about the unscientific way in which we forecast this century’s climate: we take 29 groups of models and average them. Anyone, we repeatedly point out, who knows weather forecasting realizes that such an activity is foolhardy. Some models are better than others in certain situations, and others may perform better under different conditions. Consequently, the daily forecast is usually a blend of a subset of available models, or, perhaps (as can be the case for winter storms) only one might be relied upon.
Finally the modelling community (as represented by the football team of authors) gets it. The second sentence of the paper’s abstract says “there is now evidence that giving equal weight to each available model projection is suboptimal.”
A map of sea-surface temperature errors calculated when all the models are averaged up shows the problem writ large:
Annual sea-surface temperature error (modelled minus observed) averaged over the current family of climate models. From Eyring et al.
First, the integrated “redness” of the map appears to be a bit larger than the integrated “blueness,” which would be consistent with the oft-repeated (here) observation that the models are predicting more warming than is being observed. But, more important, the biggest errors are over some of the most climatically critical places on earth.
Start with the Southern Ocean. The models have almost the entire circumpolar sea too warm, much of it off more than 1.5°C. Down around 60°S (the bottom of the map) water temperatures get down to near 0°C (because of its salinity, sea water freezes at around -2.0°C). Making errors in this range means making errors in ice formation. Further, all the moisture that lies upon Antarctica originates in this ocean, and simulating an ocean 1.5° too warm is going to inject an enormous amount of nonexistent moisture into the atmosphere, which will be precipitated over the continent in nonexistent snow.
The problem is, down there, the models are making error about massive zones of whiteness, which by their nature absorb very little solar radiation. Where it’s not white, the surface warms up quicker.
(To appreciate that, sit outside on a sunny but calm winters day, changing your khakis from light to dark, the latter being much warmer)
There are two other error fields that merit special attention: the hot blobs off the coasts of western South America and Africa. These are regions where relatively cool water upwells to the surface, driven in large part by the trade winds that blow into the earth’s thermal equator. For not-completely known reasons, these sometimes slow or even reverse, upwelling is suppressed, and the warm anomaly known as El Niño emerges (there is a similar, but much more muted version that sometimes appears off Africa).
There’s a current theory that El Niños are one mechanism that contributes to atmospheric warming, which holds that the temperature tends to jump in steps that occur after each big one. It’s not hard to see that systematically creating these conditions more persistently than they occur could put more nonexistent warming into the forecast.
Finally, to beat ever more manfully on the dead horse—averaging up all the models and making a forecast—we again note that of all the models, one, the Russian INM-CM4 has actually tracked the observed climate quite well. It is by far the best of the lot. Eyring et al. also examined the models’ independence from each other—a measure of which are (and which are not) making (or not making) the same systematic errors. And amongst the most independent, not surprisingly, is INM-CM4.
(It’s update, INM-CM5, is slowly being leaked into the literature, but we don’t have the all-important climate sensitivity figures in print yet.)
The Eyring et al. study is a step forward. It brings climate model application into the 20th century.