Quantifying uncertainty in land surface model projections under varying calibration assumptions

Computer models of physical systems are often expensive to run and have large numbers of unknown parameters, with emulators 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, including a measure of uncertainty on this prediction, and search for not implausible matches to real-world observations via history matching. We usually have many high-dimensional spatial and/or temporal fields as outputs, and we consider how to efficiently emulate and calibrate such outputs.

There are many sources of uncertainty in this procedure, and in particular when calibrating we must address the critical issue of model discrepancy (the mismatch between the real world and the model that cannot be removed by better tuning the inputs). Using simulations of the land surface model JULES, and in particular considering the uncertainty in projections of the land carbon sink under climate change scenarios, we explore the impact that different assumptions regarding model discrepancy can have on inference we make about model parameters, and on the resulting uncertainty regarding future model behaviour.

We provide an efficient emulation and calibration framework that enables modellers to input their beliefs about various land surface model outputs, and thereafter explore calibrated-model world conditional on these judgements. In particular, considering the impact such choices have on the calibration of different input parameters, identifying trade-offs and potential structural errors, and how uncertainty on the calibrated inputs propagates through to uncertainty on projections of the carbon sink up to 2100.