The dynamics of weather and climate encompass a wide range of spatial and temporal scales which are coupled through the nonlinear nature of the governing equations of motion. A stochastic climate model resolves only a limited number of large-scale, low-frequency modes; the effect of unresolved scales and processes onto the resolved modes is accounted for by stochastic terms. Here, such low-order stochastic models are derived empirically from time series of the system using statistical parameter estimation techniques.
The first part of the talk deals with subgrid-scale parametrisation in atmospheric models. By combining a clustering algorithm with local regression fitting a stochastic closure model is obtained which is conditional on the state of the resolved variables. The method is illustrated on the Lorenz '96 system and then applied to a model of atmospheric low-frequency variability based on empirical orthogonal functions.
The second part of the talk is concerned with deriving simple dynamical models of glacial millennial-scale climate variability from ice-core records. Firstly, stochastically driven motion in a potential is adopted. The shape of the potential and the noise level are estimated from ice-core data using a nonlinear Kalman filter. Secondly, a mixture of linear stochastic processes conditional on the state of the system is used to model ice-core time series.
view more