A few weeks ago I was at a meeting in Cambridge that discussed how (or whether) paleo-climate information can reduce the known uncertainties in future climate simulations.
The uncertainties in the impacts of rising greenhouse gases on multiple systems are significant: the potential impact on ENSO or the overturning circulation in the North Atlantic, probable feedbacks on atmospheric composition (CO2, CH4, N2O, aerosols), the predictability of decadal climate change, global climate sensitivity itself, and perhaps most importantly, what will happen to ice sheets and regional rainfall in a warming climate.
The reason why paleo-climate information may be key in these cases is because all of these climate components have changed in the past. If we can understand why and how those changes occurred then, that might inform our projections of changes in the future. Unfortunately, the simplest use of the record – just going back to a point that had similar conditions to what we expect for the future – doesn’t work very well because there are no good analogs for the perturbations we are making. The world has never before seen such a rapid rise in greenhouse gases with the present-day configuration of the continents and with large amounts of polar ice. So more sophisticated approaches must be developed and this meeting was devoted to examining them.
The first point that can be made is a simple one. If something happened in the past, that means it’s possible! Thus evidence for past climate changes in ENSO, ice sheets and the carbon cycle (for instance) demonstrate quite clearly that these systems are indeed sensitive to external changes. Therefore, assuming that they can’t change in the future would be foolish. This is basic, but not really useful in a practical sense.
All future projections rely on models of some sort. Dominant in the climate issue are the large scale ocean-atmosphere GCMs that were discussed extensively in the latest IPCC report, but other kinds of simpler or more specialised or more conceptual models can also be used. The reason those other models are still useful is that the GCMs are not complete. That is, they do not contain all the possible interactions that we know from the paleo record and modern observations can occur. This is a second point – interactions seen in the record, say between carbon dioxide levels or dust amounts and Milankovitch forcing imply that there are mechanisms that connect them. Those mechanisms may be only imperfectly known, but the paleo-record does highlight the need to quantify these mechanisms for models to be more complete.
The third point, and possibly the most important, is that the paleo-record is useful for model evaluation. All episodes in climate history (in principle) should allow us to quantify how good the models are and how appropriate are our hypotheses for climate change in the past. It’s vital to note the connection though – models embody much data and assumptions about how climate works, but for their climate to change you need a hypothesis – like a change in the Earth’s orbit, or volcanic activity, or solar changes etc. Comparing model simulations to observational data is then a test of the two factors together. Even if the hypothesis is that a change is due to intrinsic variability, a simulation of a model to look for the magnitude of intrinsic changes (possibly due to multiple steady states or similar) is still a test both of the model and the hypothesis. If the test fails, it shows that one or other elements (or both) must be lacking or that the data may be incomplete or mis-interpreted. If it passes, then we a have a self-consistent explanation of the observed change that may, however, not be unique (but it’s a good start!).
But what is the relevance of these tests? What can a successful model of the impacts of a change in the North Atlantic overturning circulation or a shift in the Earth’s orbit really do for future projections? This is where most of the attention is being directed. The key unknown is whether the skill of a model on a paleo-climate question is correlated to the magnitude of change in a scenario. If there is no correlation – i.e. the projections of the models that do well on the paleo-climate test span the same range as the models that did badly, then nothing much has been gained. If however, one could show that the models that did best, for instance at mid-Holocene rainfall changes, systematically gave a different projection, for instance, of greater changes in the Indian Monsoon under increasing GHGs, then we would have reason to weight the different model projections to come up with a revised assessment. Similarly, if an ice sheet model can’t match the rapid melt seen during the deglaciation, then its credibility in projecting future melt rates would/should be lessened.
Unfortunately apart from a few coordinated experiments for the last glacial period and the mid-Holocene (i.e. PMIP) with models that don’t necessarily overlap with those in the AR4 archive, this database of model results and tests just doesn’t exist. Of course, individual models have looked at many various paleo-climate events ranging from the Little Ice Age to the Cretaceous, but this serves mainly as an advance scouting party to determine the lay of the land rather than a full road map. Thus we are faced with two problems – we do not yet know which paleo-climate events are likely to be most useful (though everyone has their ideas), and we do not have the databases that allow you to match the paleo simulations with the future projections.
In looking at the paleo record for useful model tests, there are two classes of problems: what happened at a specific time, or what the response is to a specific forcing or event. The first requires a full description of the different forcings at one time, the second a collection of data over many time periods associated with one forcing. An example of the first approach would be the last glacial maximum where the changes in orbit, greenhouse gases, dust, ice sheets and vegetation (at least) all need to be included. The second class is typified by looking for the response to volcanoes by lumping together all the years after big eruptions. Similar approaches could be developed in the first class for the mid-Pliocene, the 8.2 kyr event, the Eemian (last inter-glacial), early Holocene, the deglaciation, the early Eocene, the PETM, the Little Ice Age etc. and for the second class, orbital forcing, solar forcing, Dansgaard-Oeschger events, Heinrich events etc.
But there is still one element lacking. For most of these cases, our knowledge of changes at these times is fragmentary, spread over dozens to hundreds of papers and subject to multiple interpretations. In short, it’s a mess. The missing element is the work required to pull all of that together and produce a synthesis that can be easily compared to the models. That this synthesis is only rarely done underlines the difficulties involved. To be sure there are good examples – CLIMAP (and its recent update, MARGO) for the LGM ocean temperatures, the vegetation and precipitation databases for the mid-Holocene at PMIP, the spatially resolved temperature patterns over the last few hundred years from multiple proxies, etc. Each of these have been used very successfully in model-data comparisons and have been hugely influential inside and outside the paleo-community.
It may seem odd that this kind of study is not undertaken more often, but there are reasons. Most fundamentally it is because the tools and techniques required for doing good synthesis work are not the same as those for making measurements or for developing models. It could in fact be described as a new kind of science (though in essence it is not new at all) requiring, perhaps, a new kind of scientist. One who is at ease in dealing with the disparate sources of paleo-data and aware of the problems, and yet conscious of what is needed (and why) by modellers. Or additionally modellers who understand what the proxy data depends on and who can build that into the models themselves making for more direct model-data comparisons.
Should the paleo-community therefore increase the emphasis on synthesis and allocate more funds and positions accordingly? This is often a contentious issue since whenever people discuss the need for work to be done to integrate existing information, some will question whether the primacy of new data gathering is being threatened. This meeting was no exception. However, I am convinced that this debate isn’t the zero sum game implied by the argument. On the contrary, synthesising the information from a highly technical field and making it useful for others outside is a fundamental part of increasing respect for the field as a whole and actually increases the size of the pot available in the long term. Yet the lack of appropriately skilled people who can gain the respect of the data gatherers and deliver the ‘value added’ products to the modellers remains a serious obstacle.
Despite the problems and the undoubted challenges in bringing paleo-data/model comparisons up to a new level, it was heartening to see these issues tackled head on. The desire to turn throwaway lines in grant applications into real science was actually quite inspiring – so much so that I should probably stop writing blog posts and get on with it.
The above condensed version of the meeting is heavily influenced by conversations and talks there, particularly with Peter Huybers, Paul Valdes, Eric Wolff and Sandy Harrison among others.