Multi-model Cross Pollination in Time via Data Assimilation

Abstract:

Nonlinear dynamical systems are frequently used to model physical processes including the fluid dynamics, weather and climate. Uncertainty in the observations makes identification of the exact state impossible for a chaotic nonlinear system, this suggests forecasts based on an ensemble of initial conditions to reflect the inescapable uncertainty in the observations. In general, when forecasting real systems the model class from which the particular model equations are drawn does not contain a process that is able to generate trajectories consistent with the data. Multi-model ensembles have become popular tools to account for uncertainties due to observational noise and structural model error in weather and climate simulation-based predictions on time scales from days to seasons and centuries. There have been some promising results suggesting that the multi-model ensemble forecasts outperform the single model forecasts. The current multi-model ensemble forecasts are focused on combining single model ensemble forecasts by means of statistical post-processing. Assuming each model is developed independently, every single model is likely to contain different local dynamical information from that of other models. Using statistical post-processing, such information is only carried by the simulations under a single model ensemble: no advantage is taken to influence simulations under the other models. A novel methodology, named Multi-model Cross Pollination in Time, is proposed for multi-model ensemble scheme with the aim of integrating the dynamical information from each individual model operationally in time. The proposed method generates model states in time via applying advanced nonlinear data assimilation scheme(s) over the multi-model forecasts. The proposed approach is demonstrated to outperform the traditional statistically post-processing in the 40-dimensional Lorenz96 flow. It is suggested that this illustration could form the basis for more general results which in turn could potentially be deployed in operational forecasting.