Why aren t climate models getting better? Bjorn Stevens, UCLA
Four Hypotheses 1. Our premise is false, models are getting better. 2. We don t know what better means. 3. It is difficult, models have rough fitness landscapes and comprise an informal hierarchy. 4. It is impossible, because the remaining imprecision in representations of the climate system are irreducible.
1. Our premise is false Text Reichler and Kim., BAMS, (2008)
Le Treut et al., IPCC AR4, (2007).
Or not... We estimate the most probable global warming for a doubling of CO2 to be near 3 C with a probable error of ±1.5 C. The role of heat uptake by ocean intermediate waters is uncertain and ``could delay the estimated warming by some decades.'' we ``cannot predict the locations and intensities of regional climate changes with confidence'' ``Existing parameterizations of cloud amounts in general circulation models are physically very crude... It must thus be emphasized that the modeling of clouds is one of the weakest links in the general circulation modeling efforts'' Charney et al., NRC (1979)
Clouds and Climate Change GFDL AM2 clouds act to enhance the warming (positive effect) clouds act to mitigate the warming (negative effect) positive cloud effect, larger climate sensitivity
Clouds and Climate Change GFDL AM2 clouds act to enhance the warming (positive effect) clouds act to mitigate the warming (negative effect) positive cloud effect, larger climate sensitivity NCAR CAM3 negative cloud effect, smaller climate sensitivity
Clouds as ultimate, rather than proximate, sources of bias GFDL AM2 NCAR CAM3
Clouds as ultimate, rather than proximate, sources of bias... the modelling of time dependent clouds is perhaps the weakest aspect of the existing general circulation models and may be the most difficult task in constructing any reliable climate model --- Arakawa (WMO,1975) It must thus be emphasized that the modeling of clouds is one of the weakest links in the general circulation modeling efforts --- Charney (NRC,1979) The physical process contributing the greatest uncertainty to [feedbacks] on this time scale [10-100 years] appears to be clouds --- Hansen et al., (Geophys Monographs 1984) GFDL AM2 NCAR CAM3 Probably the greatest uncertainty in future projections of climate arises from clouds and their interactions with radiation... even the sign of this feedback remains unknown --- IPCC (TAR 2001) Cloud effects remain the largest source of uncertainty in model based estimates of climate sensitivity --- IPCC (AR4 2007)
2. We don t know what better means In NWP it is possible to trace the quantitative increase in skill over time because forecasts have been evaluated against observations in a consistent manner for decades... climate models used to make longterm projections have not been subject to uniform assessment over time. (Pincus et al., 2008). In climate modelling, there is a subject[ive] judgement of what errors to minimize. The measures of success and failure are much more vague and personal. (Williamson, 2002).... but because the development of robust metrics is still at an early stage, the model evaluations presented in this chapter are based primarily on experience and physical reasoning, as has been the norm in the past. (Randall et al., 2007). Several important issues complicate the model validation process. First,... it [is] difficult to decide which of the many aspects of climate are important for a good simulation. Second, climate models must be compared against present (e.g., 1979 99) or past climate,... [neither of which is] an independent dataset since it has already been used for the model development. Third, there is a lack of reliable and consistent observations for present climate... Finally, good model performance evaluated from the present climate... does not necessarily guarantee reliable predictions of future climate. (Reichler and Kim 2008).
Some possible reasons We don t agree on which problem we are trying to solve. Key observations are unreliable or unavailable (liquid water). The lack of proxies for the climate sensitivity in the current simulation record reflects model tuning. For instance, most models reasonably represent the top of the atmosphere energy budget for demonstrably wrong distributions of clouds (implicit flux adjustment). In many instances climate modeling lacks an institutional framework.
3. It is difficult (rough fitness landscape*) *Held, BAMS (2005, early draft thereof)
4. It is impossible (models are not structurally stable)* p(x y) there thus emerges a clear imperative to better understand the consequences of the choices that are commonly made in constructing a plausible AOS model... But? p(x y ) The analysis of multimodel seasonal forecast reliability diagrams... provides a means to quantitatively discount the climate change probabilities in the light of diagnosed unreliability (Palmer et al., 2008). x * McWilliams, PNAS (2007)
Evolution of forecast skill for northern and southern hemispheres But NWP has made steady progress... data, resolution, physics Improvement of NWP 8 August 2001 Simmons and Hollingsworth, QJRMS (2002)
Resolution Points around the Equator Number of Levels Year Year
Physics Here we compare the age as measured by publication date of a particular scheme minus the publication date of the model as a whole for six models and five major parameterizations: (i) gravity wave drag; (ii) deep convection; (iii) shallow convection; (iv) planetary boundary layer; (v) cloud scheme. Parameterization Age Guess which ones are forecast models. Model Index
Inferences 1. Progress has been made in making more complex models work, i.e., progress has principally rested on our ability to expand the problem. 2. More complex models are increasingly non-fundamental (no convergence limit), but guarantee new results (managers like this). 3. Even so many lines of evidence suggest that more complex models inherit the problems of simpler models; thus complex models are less fundamental and yet no more reliable than simpler models; hence more emphasis should be placed on developing understanding thereby allowing us to make better basic models (scientists like this). 4. Deficiencies in models reflect deficiencies in both understanding and model craftsmanship, addressing the latter can help clarify the issues in the former. 5. Experience indicates that two independent models are on average better than a single model that is twice as expensive; hence focusing development efforts on a single model would be deleterious.
A proposal: Climate Model Hosting Center... tasked with maintaining facilities for running and archiving models provided by national centers and using these to develop climatically relevant data analyses (reanalysis), as well as measures of model performance and improvement (for instance through seasonal prediction). Such a facility could: strengthen and motivate national efforts; build on the success of PCMDI (the well recognized success of the AR4); institutionalize Palmer DEMETER (seamless prediction); archive and protect source code; provide service for national partners and external communities, alleviating basic-science institutes of current service burden; serve as a facility for hosting scientific visitors on projects of up to five years -- beneficial for countries with less cargo ; consume a lot of electricity. --- unlike NCAR it should have an espresso machine.
An alternative* idea: The abduction of Europa *Not favored by this speaker