Intelligent Earth Seminar: Laura Mansfield (University of Oxford)

4 June 14:00
Seminar Room 1a, Doctoral Training Centre, 1-4 Keble Rd

 

headshot smallsize

 

I am a Schmidt AI in Science fellow at the University of Oxford, with a broad interest in the use of artificial intelligence (AI)/machine learning (ML) and Bayesian statistics to advance climate modelling. Specifically, I develop ML subgrid-scale parameterisations for small-scale processes such as convection and atmospheric gravity waves. I am particularly interested in quantifying uncertainties associated with ML parameterisations. Previously, I was a postdoc at Stanford University, where I worked on uncertainty quantification of ML and physics-based gravity wave parameterisations, and prior to that, I completed my PhD at the University of Reading in climate model emulation and Bayesian statistics.

Uncertainty quantification for small-scale processes in climate models

 

Climate models simulate the atmospheric circulation using the governing equations, but are limited by model resolution, typically around 100km for simulations on climate timescales. Important processes occurring on lengthscales smaller than this, such as clouds, convection and atmospheric gravity waves, are not directly resolved and must instead be included through parameterisations. This involves simplifying assumptions and can add a significant source of uncertainty into climate models. Machine learning (ML) is emerging as a promising approach for learning parameterisations. ML parameterisations are typically trained on datasets generated by high resolution climate models or existing parameterisations (“offline”), but evaluated based on their performance when coupled into an existing climate model (“online”). Quantifying uncertainties associated with ML parameterisations is crucial for gaining insights into the reliability of hybrid ML-climate models.

I will discuss how we can estimate uncertainties associated with ML parameterisations, considering "epistemic" uncertainty from the ML model and "aleatoric" uncertainty originating from the training dataset. Although offline evaluation may show small uncertainties, these can propagate once coupled online, potentially leading to significant uncertainty in climate model circulation that we should consider carefully when building ML parameterisations.