Accurate Integration of Stochastic Climate Models with Application to El Niño

Abstract

Numerical models are one of the most important theoretical tools in atmospheric research, and the development of numerical techniques specifically designed to model the atmosphere has been an important discipline for many years. In recent years, stochastic numerical models have been introduced in order to investigate more fully Hasselmann's suggestion that the effect of rapidly varying ''weather'' noise on more slowly varying ''climate'' could be treated as stochastic forcing. In this article an accurate method of integrating stochastic climate models is introduced and compared with some other commonly used techniques. It is shown that particular care must be used when the size of rapid variations in the ''weather'' depends upon the ''climate.'' How the implementation of stochasticity in a numerical model can affect the detection of multiple dynamical regimes in model output is discussed. To illustrate the usefulness of the numerical schemes, three stochastic models of El Niño having different assumptions about the random forcing are generated. Each of these models reproduces by construction the observed mean and covariance structure of tropical Indo-Pacific sea surface temperature. It is shown that the skew and kurtosis of an observed time series representing El Niño is well within the distributions of these statistics expected from finite sampling. The observed trend, however, is unlikely to be explained by sampling. As always, more investigation of this issue is required.

Effects of stochastic parametrizations in the Lorenz '96 system

“…2(b) are more subtle, but again the non-stochastic and independent parametrizations yield similar probability densities, while the persistent parametrization produces a probability density that is more similar to the truth. The result that stochastic parametrizations improve the representation of the climate relative to the conventional deterministic alternative, indicates that potential problems related to numerical integration of the stochastic differential equations have not produced computational artifacts (Penland 2003;Ewald et al 2004). …”

On modelling physical systems with stochastic models: diffusion versus Lévy processes

“…As noted in equation (2.9), the correlation structure of the fast variable comes into play. Simply replacing the fast term with a Gaussian random deviate with standard deviation equal to that of the variable to be approximated, and then using deterministic numerical integration schemes, is a recipe for disaster (Sura & Penland 2002;Ewald et al 2004). …”

A Stochastic Parameterization for Deep Convection Based on Equilibrium Statistics

“…It should be noted that there are some fundamental issues in the numerical solution of stochastic differential equations, but the situation is far from hopeless (Penland 2003;Ewald et al 2004). The scheme of Buizza et al (1999) represents perhaps the simplest and best-known example of a stochastic forcing.…”