Parameterizing Cloud Fraction and Condensate using Physics Informed Machine Learning

Abstract: Cloud parameterizations are a source of major biases in global atmospheric models. In the last few years, machine learning has been shown to be able to emulate and even surpass physical parameterizations. Therefore, this study aims to develop and implement a novel machine learning-based parameterization of cloud fraction and condensate. For that purpose, observational data from CloudSat and ERA5 were aligned, integrated and processed to create a new dataset, tailored to the development of ML models. Using this dataset, a machine learning model will be trained to predict liquid and ice water content, as well as cloud fraction from temperature, humidity and other relevant meteorological variables. Results will focus on exploratory data analysis and initial performance of some ML emulators. The future plans for this study include two novel aspects. Firstly, a focus on Physics Informed Machine Learning techniques and an assessment of their benefit in terms of fidelity of the predictions, interpretability and ability to train using less data. Secondly, we will make use of Uncertainty Quantification methods and develop a non-deterministic version of the scheme whose stochastic variability will be traceable from the original data and through the ML architecture and the physics-informed elements.  

Biography: I did my bachelor’s in physics and astrophysics at the University of Birmingham, followed by a master’s in scientific computing and data analysis at Durham University. I have been working at the Met Office for almost three years, and my work consists of developing machine learning models and implementing them on our atmospheric models. I have also recently started my PhD at the University of Exeter where I will investigate physics informed ML and explainable AI techniques applied to the field of weather and climate modelling.