How dynamical models can learn from the data—an example with a simplified ENSO model

Abstract

Most dynamical models of the natural system contain a number of empirical parameters which reflect our limited understanding of the simulated system or describe unresolved subgrid-scale processes. While the parameterizations basically introduce some uncertainty to the model results, they also hold the prospect of tuning the model. In general, a deterministic tuning is related to an inversion of the model which is often impossible or requires considerable computing effort for most climate models. Another way to adjust the model parameters to a specific observed process is stochastic fitting where a set of parameters and model output are taken as random variables. Here, we present a dynamical–statistical approach with a simplified model of the El Nino–Southern Oscillation (ENSO) cycle whose parameters are adjusted to simulated and observed data by means of Bayesian statistics. As ENSO model, we employ the Schop–Suarez delay oscillator model. Monte Carlo experiments highlight the large sensitivity of the model results to varied model parameters and initial values. The statistical adjustment is done by Bayesian model averaging of the Monte Carlo experiments. Applying the method to simulated data, the posterior ensemble mean is much closer to the reference data than the prior ensemble mean. The learning effect of the model is evident in the leading empirical orthogonal functions and statistically significant in the mean state. When the method is applied to the observed ENSO time series, the ENSO model in its classical setup is not able to account for the temporally varying periodicity of the observed ENSO phenomenon. An improved setup with continuous adjustment periods and extended parameter range is developed in order to allow the model to learn from the data gradually. The improved setup leads to promising results during the twentieth century and even a weak forecast skill over 6 months. Thus, the described method offers a promising tool for data assimilation in dynamical weather and climate models. However, the simplified ENSO model is barely appropriate for operational ENSO forecasts owing to its limited physical complexity.