Reducing uncertainty in land surface model projections

It is uncertain how much CO2 will be taken up by the land under climate change, with a wide range of possible futures captured across CMIP6 (Coupled Model Intercomparison Project) models. Perturbing parameters within the UK Land Surface Model, JULES, suggests this uncertainty may be even wider, and it is unclear whether the land surface will overall be a carbon source or sink under different futures. Better constraints of the uncertainty of this part of climate models is key for reducing uncertainty in climate projections.

When running simulations with JULES, there are many drivers of output uncertainty. Model simulations can be computationally expensive, and statistical or machine learning emulators are commonly used as an approximation for the true model, allowing the high-dimensional input and output spaces to be more fully explored, and enabling calibration of inputs to be more feasible. This project will build on past work emulating JULES, in particular for projections of land carbon uptake under different climate change scenarios.

The overarching aims of this project are to reduce uncertainty in simulations of JULES under future emissions scenarios, to better understand whether the carbon sink is negative/positive, and the implications this has for future climate. As an efficient way of exploring these aims, the project will train ‘emulators’ for exploring the model output, for use in calibration to observational data, and for future projections. The project will combine emulation and calibration, but within each of these areas there are many potential avenues:

– Emulation: across large number of unknown input parameters, across boundary conditions, across future emissions scenarios, or across all these unknowns at once. Training emulators in large input spaces, and across multiple related time series outputs is non-trivial and may require novel methodology, building on existing statistical and machine learning approaches. Training emulators to coupled versions of the model is challenging due to expense, but could be achievable via multi-level emulation techniques.

– Calibration: history matching is a powerful technique to identify ‘good’ versions of model inputs, but which outputs we emulate and calibrate to, and how we define structural error (differences between model-world and reality that can’t be removed by tuning) will affect future projections, and work is required to properly account for this. Within this, the project could gain an understanding in how sensitive future projections are to inputs and the assumptions we make during calibration.

Regardless of the route taken, this will feed into better understanding and constraint of uncertainties relating to future projections of the carbon cycle, whilst advances in methodology will be more widely applicable to other environmental problems.