Abstract

Computer models of physical systems are often expensive to run, resulting in a small number of model evaluations in the large input space, therefore emulators are trained on the model output for use as a cheap approximation of the true model. Using an emulator, we can efficiently predict the model output at unseen inputs, and search for plausible matches to real-world observations. In environmental applications, we usually have many high-dimensional spatial and/or temporal fields as outputs. We consider how to efficiently emulate and calibrate such outputs, whilst addressing the critical issue of model discrepancy (the mismatch between the real world and the model that cannot be removed by better tuning the inputs), considering alternative frameworks for learning about this term via reframing the calibration problem in order to avoid the need for strong prior judgements on its high-dimensional form. We demonstrate the emulation and calibration procedure for a land surface model, addressing uncertainty in projections of the land carbon sink under climate change.