Vassili Kitsios: Physics-constrained climate emulator of the coupled global atmosphere, land, ocean and sea-ice Earth system
Introduction: The world is grappling with how to navigate the path to a lower carbon economy, that respects the Earth’s finite resources, whilst still meeting human requirements (e.g. food and energy security). This requires an assessment of physical climate risks for a broad range of future economic, and hence emissions, scenarios. Due to their computational expense, general circulation models (GCMs) simulate a limited number of such pathways. Here we employ data-driven methods to develop emulators that rapidly estimate the current and future climate.
Methodology: A reduced-order model (ROM) of the coupled global atmosphere, land, ocean and sea-ice system is developed by projecting the equations of motion onto an orthogonal basis of multi-variate three-dimensional spatial patterns (or modes). This transforms the spatio-temporal partial differential equations, into ordinary differential equations dependent upon only time and mode index. The modes are learnt from climate reanalyses. The ROM coefficients are calculated using a regression approach with additional hard constraints promoting energy conservation. Remaining model errors are accounted for via stochastic parameterisation and generative machine learning approaches.
Results: Using this approach, the ROM has reproduced statistical properties of the underlying historical data, and generated plausible future climates when forced by prescribed future radiative forcings.
Audience: This approach can evaluate many more potential future emissions scenarios, than one could achieve using GCMs alone. This enables stakeholders to make data-informed decisions concerning their risk exposure during potential future economic transitions.
Keywords: Climate Emulator, Atmosphere, Ocean, Land, Sea-ice, Model Reduction, Machine Learning, Stochastic Parameterisation.