Rather than relying on the outcome of a single climate model, we run ensembles of thousands of models. Each version of the model within an ensemble is very slightly different from the others but still plausibly represents the real world, and so produces slightly different outcomes. We can then look at outputs, such as the average, over the entire ensemble to give us our modelling results. This allows us to understand the uncertainty in the models better.

Getting volunteers to run members of an ensemble on their home computer has allowed CPDN (short for climateprediction.net) to create much larger climate model ensembles than has been possible before.

There are several ways of creating the small variations between the models within an ensemble, each designed to look at a different question about the model:

• Initial Condition Ensembles
• Perturbed Physics Ensembles
• Forcings Ensembles

Initial Condition Ensembles

Initial condition ensembles involve the same model, with the same atmospheric physics, run from a variety of different start dates. Because the climate system is chaotic, tiny changes in things such as temperatures, winds, and humidity in one place can lead to very different paths for the system as a whole. We can work around this by setting off several runs started with slightly different starting conditions, and then look at the evolution of the group as a whole. This is similar to what they do in weather forecasting.

Initial condition ensembles allow us to investigate the internal variability of the climate system. Such ensembles, driven with present day forcing conditions, can be used to validate a climate model by comparing the model output with actual observed data. This process is called a hindcast: it’s like a forecast, only you know the outcome.

We know what happened in the historic period we simulate because we have observed data, but it’s still a challenge for the model to do a good job of simulating it. In a ‘good’ model we expect initial condition ensembles to represent weather events with a similar frequency of occurrence as that in the observed records. Only ‘good’ models will be used for predicting the future. This does not necessarily mean they are actually good for that job too, but it does at least rule out models that are ‘bad’ at simulating the past and so won’t produce reliable future projections.

Initial condition ensembles also let us investigate how sensitive certain processes in the climate system are to changing initial conditions.

Every single model run by participants will be unique.

Perturbed Physics Ensembles

Perturbed physics ensembles form a major scientific focus of the whole project, especially using the coupled model. Modern climate models do a good job of simulating many large-scale features of present-day climate. However, these models contain large numbers of adjustable parameters which are known, individually, to have a significant impact on simulated climate. While many of these are well constrained by observations, there are many which do not directly relate to observed quantities and are subject to considerable uncertainty as they are obtained in a trial-and-error way, trying to best match observations.

We do not know the extent to which different choices of parameter-settings may provide equally realistic simulations of 20th century climate but different forecasts for the 21st century. The most thorough way to investigate this uncertainty is to run a massive ensemble experiment in which each relevant parameter combination is investigated. Thus the perturbed physics ensemble is a central feature of the CPDN project.

The knowledge we gain from this experiment about model sensitivity will enable the scientific community to design better models in the future. By perturbing parameters which control the model’s physical processes (such as cloud formation) it is also possible to see different realisations of climate change.

As in the initial condition experiment everybody’s model will be unique because each will have a different combination of parameters.

Forcings

Forcings are the things which drive the climate system. We call them forcings because they force the system from the outside: if these things change, we expect the climate system to respond.

For instance, if the sun puts out more energy, we would expect the Earth to heat up. Examples of forcings include solar variability, sulphates from volcanic activity and greenhouse gases. These are all treated as external to the climate system whereas the chaotic variability we target in the initial condition ensemble is due to factors internal to the climate system.

We normally distinguish between natural forcings, such as volcanoes and solar activity, and human-induced forcings, such as greenhouse gases, aerosols or land use change.

The first experiment we normally run in a new project is one with present day forcings to investigate whether or not the model responds to the measured forcings in a similar way to the real climate system.

In most projects we want to analyse the sensitivity of the model to different forcing conditions. In the geoengineering projects, for example, we change the amount of sulphate aerosols in the aerosol forcing to a higher concentration, aiming to test how the climate system would react to such a measure.

In the projects that aim to identify the influence of anthropogenic forcing, mainly greenhouse gas forcing, on present day climate, we change the greenhouse gas forcing files to simulate a world that might have been without anthropogenic climate change. Most Weather@Home projects comprise this experiment.

All projects aiming at future projections need to include forcings for the future, which we obviously cannot measure but have to assume. Therefore we need a future forcings ensemble with variations of different forcings because we don’t know what the sun or the volcanoes are going to do over the next fifty years. We also don’t know how levels of greenhouse gases are going to change over that period. So we’re going to run a large number of different futures which seem to be plausible, in which we vary solar, sulphate and greenhouse forcing, to span what we hope will be the likely range.

In a lot of future experiments, however, we are interested in isolating the effect of a certain forcing, such as a doubling of the amount of carbon dioxide in the atmosphere. In such an experiment we keep solar and volcanic forcings constant, i.e. we assume there will be no major volcanic eruption in the future period of interest, and only change the greenhouse gas forcing. Applying this strategy it is possible to identify the influence of a single type of forcing if comparing the future simulations with present day simulations.