| Summary: | The prediction of the Earth's climate system is of immediate importance to many decision-makers. Anthropogenic climate change is a key area of public policy and will likely have widespread impacts across the world over the 21st Century. Understanding potential climate changes, and their magnitudes, is important for effective decision making. The principal tools used to provide such climate predictions are physical models, some of the largest and most complex models ever built. Evaluation of state-of-the-art climate models is vital to understanding our ability to make statements about future climate. This Thesis presents a framework for the analysis of climate models in light of their inherent uncertainties and principles of statistical good practice. The assessment of uncertainties in model predictions to-date is incomplete and warrants more attention that it has previously received. This Thesis aims to motivate a more thorough investigation of climate models as fit for use in decision-support. The behaviour of climate models is explored using data from the largest ever climate modelling experiment, the climateprediction.net project. The availability of a large set of simulations allows novel methods of analysis for the exploration of the uncertainties present in climate simulations. It is shown that climate models are capable of producing very different behaviour and that the associated uncertainties can be large. Whilst no results are found that cast doubt on the hypothesis that greenhouse gases are a significant driver of climate change, the range of behaviour shown in the climateprediction.net data set has implications for our ability to predict future climate and for the interpretation of climate model output. It is argued that uncertainties should be explored and communicated to users of climate predictions in such a way that decision-makers are aware of the relative robustness of climate model output.
|
|---|
