Tuning del bilancio radiativo di un modello climatico a complessità intermedia utilizzando emulatori statistici

The climate system is a complex dynamical system, consisting of different major components and the interactions between them and numerical climate models are an important tool for studying its dynamics, climate change and the main processes at work. Climate model tuning (or calibration) is a standard practice in the field, also explored in the literature, consisting in the adjustment of a set of poorly constrained parameters of empirical formulas, or parameterizations, implemented to account for the large-scale influence of sub-scale processes, in order to improve the agreement between the simulation output and the observations. Model tuning is often performed manually by changing parameters one at the time and in order to improve on this arbitrary technique, different automatic and semi-automatic methods have been proposed in literature. A significant problem is given by the large computational cost of performing enough experiments to sufficiently explore the parameter space. An interesting solution is provided by the use of statistical emulators, or metamodels, that use a small set of points to build statistical models able to predict the output of the climate model. In this work, I explore this approach applied to a climate model of intermediate complexity (EMIC), together with the optimization of a cost function to determine optimal sets of parameters. The considered EMIC is the Planet Simulator (PLASIM), that, thanks to its low computational cost allows to study some of the open problems in climate model tuning, in particular the best sampling method to sample the parameter space and the best trade-off between the size of the training set and complexity of the emulator. The possibility of introducing a penalization of the parameters displacements from their default values to prevent the convergence at the border of their feasible range is investigated, as well as the benefit of implementing a relative emulator rather than an absolute one. Moreover, the possibility of using an emulator to tune the model at a higher resolution is studied. The impact of the recommended parameter changes on the model climatology is analyzed and the resulting model mean state for a range of parameters is compared with available observations.