Model Bending: Towards Dealing with Model Inadequacies in Data Assimilation and Forecasting Using a Single Model Structure

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

There are two major sources of errors that account for inaccurate weather predictions from numerical models: the initial-condition errors and model inadequacies. In this project, the PIs will tackle the model inadequacies problem through data assimilation and adjustment of physical parameters during forecast period. They will use the model output statistics (MOS) produced from forecasts as pseudo-data of the system's future states in a four dimensional variational data assimilation (4d-Var) to find the initial conditions that minimize the mismatch between the MOS forecasts and the model forecasts. A second, independent, application of MOS is then applied to the resulting forecast 4d-Var states. This procedure is termed as a "forecast 4d-Var" as it takes into account the model errors through 4d-Var during forecast. The second part of the research is to dynamically alter model parameters, instead of the initial conditions, using 4d-Var to "bend" the model toward the true state of the atmosphere. Self-correcting models will be developed to allow selected parameters to be slowly varying functions of time and space based on physical constraints. During this period, MOS for these parameters are obtained and then used to "predict" optimized parameters in forward runs. This procedure is termed as "model bending". Models of varying levels of complexity will be used to gain a better understanding of model inadequacies overall. The research will bring more attention into model errors in numerical predictions. It has the potential of improving operational forecasts and directing future model development. If successful, it would lead to a major breakthrough in numerical modeling. The project provides good opportunities for training graduate students in the highly needed areas of data assimilation and predictability.