Green Sahara: how? pt.2 — Model not so good

In my most recent post I outlined how atmosphere circulation models were used to develop a mechanistic understanding of how the climate system is able to would be able to produce a green Sahara in the mid-Holocene. However, there are several important issues with these modelling experiments that need to be discussed.

Modelled mid-Holocene precipitation anomaly
Reproduced from Kutzbach & Guetter (1986) 

As this image shows, the low spatial resolutions of these early models mean that one grid cell represents a large area of the continent. When combined with relatively sparse spatial coverage of the observation sites, the difficulty in the evaluating the extent of monsoon penetration into the Sahara with any great degree of precision is apparent. More detailed palaeoclimatic maps were developed over the course of time, against which models of increasing resolution could be assessed. These datasets proved key to the first phase of the Paleoclimate Model Intercomparison Project (PMIP); a collection of experiments established in the 1990s in order to evaluate the ability of eighteen different AGCMs to reproduce African Humid Period state when run with 6 ka parameters and boundary conditions. The project found considerable agreement in results between the different experiments, but that "the magnitude of the monsoon increases over northern Africa are underestimated by all the models". This reveals one of the most significant weaknesses of such complex models: the comprehensive inclusion of complex processes is restrained by the available processing power of the computers they run on. AGCMs — as the term implies — simulate atmosphere dynamics only. Other constituents of the Earth system, such as the ocean or the ice sheets, must be either prescribed or omitted.

Following the somewhat dissatisfying results of PMIP Phase 1, however, the palaeoclimate modelling community took advantage of ever-increasing computational resources to run more complex experiments, and at higher resolutions, to more success. For example, one experiment found that, when the land surface scheme is prescribed to be more representative of a vegetated Sahara, the 6 ka WAM reaches further northward and is considerably enhanced. This implies that the simulation of land surface–atmosphere feedbacks would be important in getting a more realistic monsoon, and indeed that assertion was supported by the improved results of models which included interactive soil moisture, snow or vegetation components. Nevertheless, it became apparent that the realism of these experiments was to a degree limited by the fact that they simulated atmosphere circulation only, as it was likely that the ocean also had a part to play in intensifying the mid-Holocene monsoon.

6 ka WAM precipitation reachers further northwards with a full AOGCM (dashed line) than with prescribed SSTs (dotted line). Solid line is control experiment. Reproduced from Braconnot et al (2000). 

Indeed, Kuztbach and Liu (1997) found better agreement with observations when an AGCM is asynchronously coupled to an ocean general circulation model (OGCM); by allowing the two dynamic systems to interact, atmosphere–ocean feedbacks — which had already been theorised from contemporary observations — could be simulated. With the use of this class of model it was ascertained that the strength and even direction of this feedback varied between regions; dampening monsoon rainfall in Asia while intensifying the West African Monsoon. More components could be added to provide a more nuanced understanding of the interplay between different parts of the climate system. For example, asynchronous coupling of an AOGCM to the BIOME vegetation model revealed that atmosphere-ocean feedbacks are the primary control of the flux of warm oceanic air onto land, while retention of this moisture on the continent is stabilised largely by land surface–atmosphere feedbacks. Including vegetation is therefore key to maintaining an amplified monsoon. Nevertheless, fully synchronised AOGCMs were unable to adequately simulate the rainfall conditions in the northern Sahara indicated by palaeodata, even with a dynamic vegetation and more explicit treatment of soil characteristics.

In a general sense, incorporating more components in a modelled experiment allows a wider range of climate processes to be simulated, and interact with each other, producing a stronger monsoon. However, similarly complex models can vary widely in their results, as shown in the image below reproduced from a study comparing two AOVGCMs whose atmosphere core differs.

6 ka precipitation anomaly from ECHAM and LMD models, initialised with present day (left) and green Sahara (right) vegetation. It is clear that there is much greater monsoon penetration using ECHAM.

The key takeaway from this is that the degree to which the mid-Holocene WAM is intensified by inclusion of vegetation is strongly dependent on any given model's treatment of atmospheric circulation dynamics. The same conclusion emerged at the conclusion of the second phase of the PMIP, which also found that model results vary not only in their response to mid-Holocene forcing, but also to modern forcing.

"Model biases or differences between the control experiments need to be considered to understand the response of various models"

Since much of the variation in the results the model experiments discussed here is due to uncertainty in their parameterisation of real life physics (as observed today), it is very difficult to assess how representative the modelled mechanisms are of the true climatic processes that generated the green Sahara.

None of the PMIP Phase 2 studies successfully reproduced the observed AHP conditions. It appears that thus far, even with the inclusion of oceans, vegetation, and soil, there is something in climate models which makes them too restrained in their simulations of the mid-Holocene monsoon. I'm not yet sure why. It is possible that there is some conservative bias inherent to the operation of all GCMs. Alternatively, there may still be some monsoon-intensifying mechanisms which are as-yet unknown or otherwise incompletely modelled. The next PMIP phase will involve the use of ESMs, a class of model which incorporate a far more comprehensive range of Earth System components, and therefore climatic processes, than previous generations of models.

It is worth pointing out that Moore's law may not hold much longer. I don't say this to imply that climate models will be approaching some kind of peak complexity any time soon, but rather because it's relevant to my doubts about a particular line of thinking in modelling. I'm not sure whether the best way to tackle model–data disagreement over the green Sahara is simply to chuck more computing power at the problem. That said, I don't yet have any truly valid reason to be sceptical. Either way, I think it's worth getting to the root of the issue. To extend a concept that my coursemate Damian highlighted, climate modelling involves "experience and intuition": not only to understand why models work, but also why they don't work. I guess gaining such experience is a key part of undertaking this blog.