Using machine learning to constrain the atmospheric dynamics contribution to regional climate change

"Global warming is the key metric in the public perception of climate change but regional changes, for example in weather extremes or rainfall, have a more direct impact on people's lives. These are particularly difficult to predict, however, so increasing confidence in regional impacts is arguably one of the most important challenges in present-day science. A large part of the uncertainty in regional projections arises from the complexity of atmospheric dynamics and its response to increasing atmospheric greenhouse gas concentrations.

The goal of this project is to use machine learning to build a data-driven mathematical framework for regional climate change that goes substantially beyond the simple global warming picture.

This framework will:

1) necessarily include both the thermodynamic and the dynamical response of the Earth system to greenhouse gas forcing. Here, we refer to thermodynamic mechanisms as those primarily driven by local changes in the energy budget, which is well reflected in variables such as surface temperature. Dynamical mechanisms broadly refer to shifts or modifications in the strength of the global atmospheric circulation, or changes in the remote coupling between regions of the atmosphere, which are referred to as teleconnections. Both components are intrinsically coupled due to the redistribution of energy (thermodynamics) within the Earth system as part of the circulation (dynamics).

2) put emphasis on an attempt to separate dynamical and thermodynamic drivers of regional change. Each driver will be evaluated using data from observations (e.g. satellite data) and the output of state-of-the-art climate models, i.e. sophisticated computer models used to make climate change projections.

Taking this framework as a basis, a first aim is to introduce novel metrics for regional climate change. Such metrics should be easily visualised and understood by non-experts, but better reflect the uncertainty in, and the importance of, the dynamical response including measures for extreme events such as heat waves, storms, droughts and floods. In addition, by focusing on certain world regions, the underlying physical drivers of uncertainty will be identified and tested concerning their potential to increase confidence in regional climate change projections.

The project will involve the modification and then application of machine learning algorithms to large climate datasets, for example to data from climate model simulations that is used to inform the Intergovernmental Panel on Climate Change (IPCC) and to data published by the European Centre for Medium-Range Weather Forecasts (ECMWF) or NASA. The ideal candidate should be able to demonstrate a keen interest in the physics of the Earth system and in testing out a number of different supervised and unsupervised machine learning algorithms. All coding will be carried out in Python. Good programming experience and familiarity with some machine learning packages (scikit-learn, TensorFlow etc) would be an advantage, but are not essential. Depending on the student's interests, high-resolution numerical models could be used to test the result of reducing uncertainty in coarser climate model projections on much finer spatial resolutions (e.g. county-to-city scale).
"