REE5: Uncertainty Analysis of Regional and Global Climate Forecasts

The growing world-wide debate about the consequences of climate change has highlighted the importance of understanding, assessing and improving upon the predictions of future climate.

The Earth's climate can be modelled as a nonlinear dynamical system comprising an extensive variety of physical processes in a wide range of spatial and time scales.  Hence the different methodologies used for the understanding of the climate system are inspired by nonlinear dynamics.

There are commonly three approaches in which the climate system can be studied:

• Using General Circulation Models (GCMs) to obtain numerical simulations for the past, present and future climate systems.  Typically these climate models treat the system with a deterministic approach and approximate some of the physical processes at work on a coarse-grained mesh.

• Using observation of the past and present states of the climate system can give us a rich insight on climate natural variability, which is a key factor when trying to predict the future climate.  In this field, extensive research has been carried out on time-series analysis also based on nonlinear dynamical systems.

• Using low-order models that describe either a subsystem of the climate or some aspects of the climate system. While it is unreasonable to expect solutions of lower dimensional systems to generalise to a high dimensional space such as the climate system, they provide a robust qualitative understanding of the physics and a test for GCMs, given that it is also unlikely that problems identified in the simplified models vanish in higher dimensional model versions.

Each of these methods presents limitations and thus, they cannot give us a complete picture of the climate system. Instead this project aims to combine the three approaches and obtain a better understanding on the uncertainties in future climate projections. 

On the first approach, we use output from the climateprediction.net distributed climate experiment for the formal analysis of uncertainty and the interpretation of ensembles.

The climateprediction.net experiment, in which over 200,000 participants worldwide ran coupled atmosphere/ocean models for slightly different parameter values, has generated 10 terabytes of data. These data sets are four-dimensional functions of latitude, longitude, height and time, and include information on variables such as surface and ocean temperatures, cloud cover, pressure, under the influence of human and natural climate forcing over 1920-2080. This data, stored and overseen by the e-Research centre in Oxford, has yet to be fully analysed.

So far only temperature changes from 1960-2005 have been used to constrain the predicted range of global temperature changes up to 2050. The next priority is to extend this analysis to regional changes and to variables such as rainfall prediction in the Gulf region. Oxford University is uniquely placed to make such ground-breaking and significant contributions to the study of climate change. The climateprediction.net experiment was conceived and developed in Oxford and members of the e-Research Centre are responsible for the implementation of the experiment.

We aim to quantify how recent climate observations constrain the range of validity of forecasts of future climate. Specifically we aim to use the output from the climateprediction.net experiment to map likelihood functions of the data as functions of future climates. To date only predictions of a single variable, namely temperature change, have been made up to 2080. We aim to extend this analysis to other variables of future climate, such as rainfall, using large ensembles to produce probabilistic forecasts and assess their predictions of climate change both on a global, and on a regional basis (especially for the Gulf region).

On the second approach, we aim to use observations of the past to gain understanding on climate variability.  Climate records exist on various time scales, from months to millions of years. These records indicate that climate varies in a irregular fashion on all time scales and that there are strong interactions among separated space and time scales. Thus predicting the effect of anthropogenic perturbations of radiative forcing on interdecadal variability requires a precise understanding of natural variability. Moreover, understanding natural variability is a key priority when investigating what causes transitions between different states of the climate system's behaviour. These are often called 'tipping points' in the literature.

Additionally, investigating the underlying dynamical properties of climate variability it is key for assessing future climate projections because the state of the art climate models cannot encompass all temporal and spatial scales nor do they include all possible components and processes.  GCMs model runs often exhibit significantly lower natural variability and in general concentrate on modelling a particular frequency band. Variables that evolve on slower timescales can be modelled as quasi-adiabatic parameters and variables evolving more rapidly can be modelled as random fluctuations.

Therefore, when looking at climatic records we should ask what is the information content and the frequency band of a model, and which features of such a record are the ones a model of behaviour in that band should capture. In this context we can evaluate climate models and constraint the reliability of their future climate projections.

In particular, we shall use climateprediction.net perturbed physics ensemble to investigate correlations between different time scales.  In this manner, we can evaluate model runs by using metrics inspired by time series analysis.

Finally, we aim to collaborate in the analysis of a low-order model for the dynamics of the sea ice, as a subsystem of the climate. This allows us to identify mathematically and physically rigorous characteristics of the physical interactions in the system, and so gain the ability to analyse specific bifurcation phenomena that GCMs have not been able to reproduce reliably with the current state of the art generation models.