ABSTRACT
Extreme precipitation events in the midlatitudes are often driven by large-scale atmospheric circulation patterns, which can provide valuable sources of predictability from days to months ahead. In this work, we develop a targeted dimensionality reduction approach based on variational autoencoders to identify dynamical drivers that are informative of a local target variable and predictable at subseasonal to seasonal leadtimes. We apply the method to identify circulation regimes targeted to extreme precipitation events over Morocco, and study their subseasonal predictability as well as their modulation by subseasonal teleconnections from the stratosphere and Madden-Julian Oscillation. We further show that the approach can be extended to study multiple interacting teleconnections at seasonal timescales by embedding a directed acyclic graph in the latent space of the variational autoencoder.
BIO
Fiona is a PhD student working with Prof Ted Shepherd at the University of Reading and Prof Marlene Kretschmer at the University of Leipzig on developing causal machine learning methods to study teleconnections in the climate system. She also developed the open-source software package ibicus for the implementation and evaluation of statistical bias adjustment of climate models and was a visiting researcher at the Alan Turing Institute in London during her PhD. Fiona holds an MSc in Mathematical Physics (University of Edinburgh) and an MSc in Environmental Change and Management (University of Oxford). Prior to starting her PhD, she worked for two years at a not-for-profit organisation on climate finance and was part of the international Mercator Fellowship programme, working on finance for climate resilience and loss and damage.
