Multi-model ensembles in Climate Modelling

Climate change is now widely recognised as a critical issue, frequently mentioned in news stories both with reference to its current impact (such as observed or perceived changes in yearly weather patterns) and its potential future effect on our physical environment. However, different climate simulators can produce models with quite different outputs. This modelling uncertainty is well documented and has for some time been leading toward a trend of using multi-model ensembles, whereby a number of models are used to make predictions based on the observed data, and their collective predictions are aggregated in some way. The result thus obtained should reduce the bias of each individual model, giving a more accurate - and therefore more useful - overall prediction. This approach is outlined in Professor Chandler's 2013 paper, in which a Bayesian analysis is applied to a multi-model ensemble to obtain refined posterior estimates of the quantities of interest. Having discussed the subject with Professor Chandler, it seems that there is scope to continue and expand this work.

I would like to propose two potential research projects in this area; the exact nature of the project undertaken will be decided after further discussion of the data available, and according to the most pressing need:
1. Develop a multi-model ensemble for a particular application, evaluating the potential impact of climate change in the study area and assessing the performance and sensitivity of the generated model.
2. Extend the multivariate Normal Bayesian analysis applied in the 2013 paper to another framework: taking either a dynamic linear modelling approach, or a nonparametric Bayesian approach, to predicting future values of the variables of interest.

The methodology required will vary depending on the research project selected. In the first, I would expect to draw on a range of established climate models, 'training' them on the observed data available and using them to predict how the variables of interest (for example, precipitation or temperature) are likely to change under any given scenario. The outputs from these disparate models will then be combined, and an overall prediction for each variable of interest obtained by performing Bayesian inference, assuming each of the variables to have a Normal distribution. This will provide information on how certain key aspects of the climate might change, given the data observed to date and our modelling assumptions, and so can be used to assist in planning how best to mitigate or manage the effects.
In the second project, a multi-model ensemble will still need to be generated based on a set of climate observations, but instead of using a Bayesian approach to perform inference, an alternative method will be applied; either a Dynamic Linear Model will be fitted, in which the posterior distribution of the variables is updated sequentially as new values of the variables are observed, or nonparametric Bayesian inference will be used, without assuming that the variables have any particular distribution. The theoretical properties of the chosen method will be described and evaluated, along with the accuracy and sensitivity of its predictions, so that other researchers can choose whether to adopt this or a similar approach in their own work.