Using statistical emulators in the evaluation and constraint of uncertainty in a global aerosol model
Speaker: Dr Jill Johnson (University of Leeds)
Venue: MANDEC Lecture Theatre, 3rd Floor, University Dental Hospital, Higher Cambridge Street, Manchester, M15 6FH
Abstract: Aerosols, tiny particles in the atmosphere from natural and pollution sources, affect the amount of radiation reflected back into space both directly and by affecting the reflective properties of clouds. It is important to know how much radiation aerosols reflect, and how this has changed over time, as it helps us to understand the Earth’s climate and energy balance and the extent to which humans are affecting it.
Climate models can be used to simulate the global distribution of aerosols and predict how the climate has responded to changes in the concentrations of these particles over time (the aerosol radiative forcing): from the pre-industrial (1850) to the present-day. This cannot be quantified using observations directly as we cannot go back to observe the past. However, climate models are extremely computationally intensive to run and such predictions are very uncertain as the models have many inputs (parameters), the values of which are uncertain. At Leeds, we have generated two large perturbed parameter ensembles (PPEs) of model simulations from the UK Met Office HadGEM3-GA6-UKCA (vn8.4) model (the aerosol component of the UK Earth System model) in which we explore the effects of many uncertain aerosol and atmospheric model parameters. By using these PPEs to construct Gaussian process emulators (surrogate models) of different aerosol and radiative properties that output from the climate model, we are able to densely sample and evaluate the model uncertainty in each of these properties as well as in the predicted aerosol forcing. We then confront and constrain this uncertainty with a diverse set of present-day aerosol observations.
In this presentation I will describe our approach for evaluating and constraining the model uncertainty and outline some challenges that arise in the implementation of this process, which include how we account for heterogeneity in the observations and the effects of model equifinality. I will present some recent results and show that some constraint on the uncertainty in aerosol forcing is achievable when we use the observations to reduce a large sample of model variants (over 1 million) that cover the full parametric uncertainty to less than 5% that are observationally plausible.
Bio: Jill Johnson is an applied statistician working as a research associate in the aerosol research group at the Institute for Climate and Atmospheric Science, University of Leeds. Her research interests lie in the application of statistical methods to explore the behaviour of complex environmental processes and systems, with a focus on modelling extreme events and uncertainty quantification in complex models.
Her work involves using statistical experiment design techniques, surrogate modelling (emulation) approaches and sensitivity analysis to quantify the important uncertainties, and history matching techniques to compare the models against observations. She focuses on the quantification and constraint of key uncertainties in complex aerosol and cloud models.