Using Optimisation Algorithms to tune Climate Models (OptClim)

Lead Research Organisation: University of Edinburgh
Department Name: Sch of Geosciences

Abstract

OptCliM will bring into climate modelling advances from mathematical optimization research. Our focus is upon parameterised processes that represent physics that are unresolved within climate models. These unresolved processes are represented through equations that include fixed parameters, with a typical climate model having around a hundred parameters. For example, thunderstorms not only generate heavy rain but are also one route for moisture into the atmosphere. One of the parameters expresses the rate at which moist air in the storm is mixed into the atmosphere. A range of values for each parameter is consistent with theory and measurement with changes in some parameters having a dramatic effect on future climate predictions. It is therefore necessary to have realistic parameter values in order to adequately model past or future climates.

OptCliM responds to the need for an automatic and objective method to produce models consistent with reality. Currently the values used in climate models are chosen by manually adjusting several of them until the model produces an acceptable simulation of the current average climate. This process is very expensive in person time; it is not objective, not reproducible, and relies heavily on individual, if expert, judgement.

OptCliM will develop iterative methods that use optimisation algorithms to automatically adjust many parameters so that models are consistent with observations. Beginning from any set of parameter values within the allowed ranges, the optimisation algorithm determines an initial set of model configurations to be run. On completion of these runs, the simulations are compared against the observations, and used to define parameter values for further runs until progress halts or the difference between simulation and observations are small. The challenges in applying such methods to climate models arise from the inherent noisiness of climate, and the computational expense of each model run. We will bring into climate modelling three alternative algorithms to find which is most effective in terms of making a model consistent with a range of different observations, and achieving that goal with minimum computing time and cost.

OptCliM will:
1) Allow researchers to more easily generate parameter sets that produce realistic models allowing a better understanding of past and future climate change.
2) Provide an objective and transparent method to combine models and specified observations.
3) Through our impact plan contribute to the development of the new UK earth system model, UKESM1.
4) Open further development of methods for a more systematic exploration of uncertainty in climate modelling, for example generating parameter value sets that sample observational uncertainty to lead to a cloud of plausible models.

Planned Impact

Our project aims to trial out methods using an earlier generation of climate models but having done that apply them to the US Dept of Energy Climate models to see how robust our findings are. Our impact plan goes beyond this by focusing on two applications of our techniques; first to use our techniques for the development of the joint NERC/Met Office Earth System Model. We aim to work with the Met Office to apply our techniques so transferring knowledge of the approach to them and involving them in its evaluation. Second, to apply one of our techniques to a different field of modelling - that of micro-magnetic modelling, to translate the approaches we find work in climate modelling to other domains.
All three groups have agreed to work with us and provide computational resources to support the trials we plan.

Publications

10 25 50
 
Description That optimisation methods work and can generate stable coupled climate models. We have explored up to 14 parameters. Some of the cases have been ran as fully coupled ocean/atmosphere climate models. Their response to CO2 increases have been explored and we have found that uncertainties in response are small. This suggests that perturbed physics uncertainties are small if the model is calibrated to observations.
Exploitation Route More papers to be written and community to be persuaded approach worth using.
Sectors Environment