Ensemble prediction has become an indispensable tool in weather forecasting. One of the issues in ensemble prediction is that, regardless of the method, the prediction error does not map well to the underlying physics (in other words, error estimates do not project strongly onto physical structures). The main objective is to show, through a series of experiments, evidence that a nonlinear approach will allow us to isolate and eliminate a deterministic portion of the error in an ensemble prediction. The experimental basis is provided from results obtained from three increasingly sophisticated chaotic systems, which incorporate atmospheric transitions and hemispheric structure. Using neural networks to probe the deterministic component of forecast error, it can be shown that the error recovery relates to the underlying type of flow and that it can be used to forecast transitions in the atmospheric flow using ensemble data. This research will expand on these findings to consider increasingly sophisticated atmospheric models including a full primitive equation ensemble modeling system.
Intellectual merit: Completion of this work result in more accurate prediction of transitions in the large-scale atmospheric flow and precisely relate the error to the underlying state, thereby gaining insight into the predictability of such transitions.
Broader impacts: The research will provide training to a graduate student in neural networks and diagnostic analysis of weather prediction errors. The results will aid the broader weather prediction research community and will be of potential benefit for societal and policy considerations.
