Developing a multi-level Gaussian process emulator of an Atmospheric General Circulation Model for palaeoclimate modelling

The study of past climates provides a unique opportunity to test our understanding of the Earth system and our confidence in climate models. The nature of this subject requires a fine balance between complexity and efficiency. While comprehensive models can capture the system’s behaviour more realistically, fast but less accurate models are capable of integrating on the long timescale associated with Palaeoclimatology. In this thesis, a statistical approach is proposed to address the limitation of our simple atmospheric module in simulating glacial climates by incorporating a statistical surrogate of a general circulation model of the atmosphere into our Earth system modelling framework, GENIE. To utilise the available model spectrum of different complexities, a multi-level Gaussian Process (GP) emulation technique is proposed to established the link between a computationally expensive atmospheric model, PLASIM (Planet Simulator), and a cheaper model, EMBM (energy-moisture balance model). The method is first demonstrated by emulating a scalar summary quantity. A dimensional reduction technique is then introduced, allowing the high-dimensional model outputs to be emulated as functions of high-dimensional boundary forcing inputs. Even though the two atmospheric models chosen are structurally unrelated, GP emulators of PLASIM atmospheric variables are successfully constructed using EMBM as a fast approximation. With the extra information gained from the cheap model, the emulators of PLASIM’s 2-D surface output fields, are built at a reduced computational cost. The emulated quantities are validated against simulated values, showing that the ensemble-wide behaviour of the spatial fields is well captured. Finally, the emulator of PLASIM’s wind field is incorporated into GENIE, providing an interactive statistical wind field which responds to changes in the bound- ary condition described by the ocean module. While exhibiting certain limitation due to the structural bias in PLASIM’s wind, the new hybrid model introduces additional variations to the over-diffusive spatial outputs of EMBM without incurring a substantial computational cost.