University of Southampton - National Oceanography Centre
Large complex computer models are necessary if we are to make predictions about the climate in the longer term. Such models are expensive to run so we cannot use naÃ¯ve Monte Carlo methods for example for inference. To avoid this problem of needing to run the model too often we use emulators, or surrogate models. Basically we build a statistical model of the deterministic dynamical climate simulator. Using the risk of the collapse of the Atlantic overturning circulation as an illustration I will outline how we analyze such models. First we need to design a computer experiment where every run can be very expensive. Current designs will be outlined and I will present some new results on sequential designs. This training set of model runs is then used to build a Gaussian process emulator. I will explain how we verify that we have a good emulator and how we can use the emulators to make inferences about model behaviour and the risk of future climate events.