RAPID-RAPIT: What is the risk of thermohaline circulation collapse?

The aim of this experiment is to assess the risk of the Atlantic meridional overturning circulation (AMOC) collapse in the coming century. The models issued to volunteers for this experiment are identified by the inclusion of HADCM3N in the model name.


The Atlantic meridional overturning circulation (AMOC), often referred to as the thermohaline circulation, plays a pivotal role in the global climate system. The AMOC transports heat northwards in the Atlantic via the Gulf Stream and North Atlantic Current, making European climate significantly warmer than it would otherwise be. Evidence from palaeoclimate records suggests that this circulation has changed dramatically in the past, and there is concern that it could be disrupted in the future. As concentrations of atmospheric greenhouse gases increase and the climate warms, it is expected that there will be increased precipitation in mid-latitudes and less formation of sea ice. This would make the surface ocean less salty, which (along with the warming) makes the surface waters less dense, and less likely to sink, meaning that the AMOC would weaken or maybe collapse completely. Previous experiments have shown that if the AMOC were to collapse, the impacts would be felt across the globe, and most severely in the North Atlantic region.

Simple models and models of intermediate complexity show that if extra freshwater is added to the North Atlantic (making the surface water less dense), then the AMOC can collapse, and even when the anomalous freshwater forcing is removed, the circulation may not return to its original state; under the same forcing, there can be two stable states for the circulation: “on” and “off”. However, when the AMOC is forced to collapse in more complicated climate models, it often recovers gradually after the forcing is removed, suggesting the off-state may not be stable. Also, when these complex models are forced by reasonably realistic future greenhouse gas projections, they tend not to show an AMOC collapse; the models used in the most recent IPCC assessment suggest that the AMOC is likely to gradually weaken over the 21st century, but not collapse abruptly.

These estimates for the response of the AMOC to future anthropogenic forcing rely on our “best guess” for many of the complex model details, and do not account for uncertainty in the model input parameters. There are many physical processes that climate models are unable to represent explicitly, and their effects must instead be parameterised. These processes include oceanic mixing and eddy activity, atmospheric convection and cloud physics. Parameterisations generally involve choosing the most suitable value for each coefficient, but it is often the case that there is a range of possible values that the coefficients can take. This means that a climate model has a multi-dimensional parameter space, with potentially billions of model versions (theoretically an infinite number), some with more realistic climates than others. Making climate projections using just one version of a model may mean that certain types of behaviour could be missed. This is particularly important for complex nonlinear systems like the AMOC. It is possible that the AMOC may be more or less stable in different parts of a model’s parameter space. To properly assess the risk of AMOC collapse, we must fully explore the possible responses.

Experiment design

This experiment uses HadCM3, a coupled model with fully dynamic atmosphere and ocean components, making it the most comprehensive climate model available on climateprediction.net (it is the higher-resolution counterpart of FAMOUS, which has been used in other CPDN experiments). It is run without flux adjustments which “nudge” the climate towards a realistic state, but have an adverse effect on important ocean processes. HadCM3 has been used extensively for climate research, and was one of the models used in the IPCC fourth assessment report. The experiment consists of a 10,000-member ensemble, covering a wide range of HadCM3’s parameter space. This is the first time that such a large ensemble has ever been carried out using this model. Ordinarily, HadCM3 has to be run on a supercomputer, meaning that it can run fast, but only a few ensemble members are possible (of order tens). With your help, running HadCM3 on climateprediction.net means that we can have a much, much larger ensemble, which is needed to properly explore the model’s capabilities.

Because of the infinite number of possible models living in parameter space, even 10,000 models would not be enough to produce valid risk assessments if taken on their own. We plan to use a state-of-the-art statistical technology called emulation in order make the maximum use of our 10,000-member ensemble. An emulator is a statistical tool that allows us to take any values of the parameters and predict, with associated uncertainty, what our climate model, HadCM3, will do at those parameter settings. The 10,000 runs of HadCM3 will be used to fit and train this statistical model.

Whilst it will take months for a complete run of HadCM3 to finish, the emulator for HadCM3 will take seconds to predict the model output. This means that once the emulator has been built by the RAPIT project statisticians, we can use it to quickly search parameter space for areas of potential AMOC collapse (and then we can run the model there to check). We can also produce a risk assessment that takes into account our uncertainty about what the model will do in any part of parameter space.
The first stage of the experiment involves spinning up each version of the model to be as close to its own equilibrium climate state as possible. Once the models are spun up, we will run them with 20th century forcings, and a variety of idealised future CO2 forcing scenarios to examine how the AMOC responds to changing CO2.

In this experiment, every ensemble member is important. It is likely that many of the model versions will have climates that look wildly unrealistic, and several of them may crash. All of this information is vital for understanding parameter space and for building useful emulators. Somewhere in parameter space there may be a region of plausible climate in which the AMOC collapses under CO2 forcing. This region might lie somewhere in between an area in which the climate looks unrealistic and an area in which the climate looks reasonable. If this is the case, then information from the extreme models will be vital for building emulators that can search for the interesting area. Please don’t be discouraged if the model you are running looks crazy – the information it tells us will be very useful.


RAPIT (Risk Assessment, Probability and Impacts Team) is a large project that is part of NERC’s RAPID-WATCH programme. The project involves scientists from the National Oceanography Centre, Durham University, University of Reading, Met Office, University of Oxford, British Antarctic Survey, Imperial College, and London School of Economics.