Testing the ability of AI-based regional models to capture hydroclimate changes of the geologic past

About the project

About the Project Background Before the Anthropocene, climate changed naturally over the course of thousands of years. However, without instrumental observations, we only can indirectly reconstruct climate from proxies in geologic archives. These show interesting changes in seasonal hydroclimate driven by subtle variations in the Earth’s orbit. Global climate models provide vital future climate projections but exhibit systematic biases when deployed on the geologic past, partly because of their coarse spatial resolution. Machine-learning offers the potential to overcome this resolution problem through ‘regional downscaling’. Yet this kind of machine-learning requires large training sets of observed weather conditions, which it then uses to extrapolate into a future with a changed climate. Simulating known past climate changes can be useful way to assess the validity of AI-downscaling and understand the uncertainties associated with it. PhD Project Orbital variations over thousands of years alter the seasonal amounts of incoming solar radiation and lead to shifts in the global monsoon that can be detected in the geological record. These climate proxy records come with uncertainties around dating and interpretation, as well as measurement errors. The first step of this project would be to formally assess the uncertainties in hydroclimate reconstructions, using a modified version of the National Physical Laboratory’s metrological framework that our group has started applied to other paleoclimate approaches. We can test the accuracy of future climate projections by exploring how well climate models capture past warm periods. The Paleoclimate Modelling Intercomparison Project (PMIP) includes several such experiments, such as simulating the climate of 127,000 years ago (during the last interglacial), and new simulations are currently underway. However, using reconstructions of past climate to rigorously discriminate between models is a challenge. You would devise a method of hydroclimate data-model evaluation that fully considers the reconstruction uncertainties as well as those arising from internal variability in the climate models. The second phase of the project would then explore the impact of increasing the spatial resolution on the ability to capture these past changes, through analysis of paired simulations performed by the USA’s National Center for Atmospheric Research (NCAR). Such fine resolution modelling requires substantial computing resources, and your newly developed evaluation method will assess whether they can be justified. Machine-learning offers the possibility of ‘downscaling’ to a finer grid for a fraction of the cost. You will deploy NVIDIA’s AI-climate model and explore whether its uncertainty estimates encompass the states captured by the NCAR simulations and reconstructions, or whether they miss a key structural uncertainty coming from extrapolation. Applicant Profile Students who excel at any sufficiently numerate subject are encouraged to apply; such as physics, mathematics, Earth sciences, geography, computer science or a related field. Prior education in climate or data science is not needed, although it would be helpful. An established interest in the topic is obviously vital, and a curiosity to learn across research disciplines. Some previous experience with scientific programming is required, and good oral and written communication skills are essential. Relevant professional experience and/or an MSc qualification would increase your competitiveness but is definitely not a requirement.

Research areas