What Adjoint Models Can Tell You About Numerical Weather Prediction: Dynamics, Forecast Uncertainty, and Impact of Observations
Space Science and Engineering Center
Cooperative Institute for Meteorological Satellite Studies
TCS Building 240
The adjoint of a numerical weather prediction model, computed as the transpose of the tangent linear approximation to the fully nonlinear model, computes the sensitivity of some aspect of the final forecast state to changes to the model state at earlier times. The adjoint is best known as an integral part of 4-dimensional variational data assimilation (4DVAR). However, by providing a concise description of how any small change to the model initial state can affect the final forecast state, the adjoint model is a powerful tool for investigating the dynamics of weather systems and how their prediction is affected by initial condition uncertainty and assimilation of observations.
The adjoint of the NASA GEOS-5 model is used in non-4DVAR application to investigate the dynamics and forecast uncertainty of the January 2000 "Surprise Snowstorm", and the adjoint of the GEOS-5 data assimilation system is used to compute the impact of observations on the intensity forecast of the 25-28 January 2015 east coast snowstorm. A sensitivity analysis of the January 2000 event provides an interesting hypothesis on what put the "surprise" in the Surprise Snowstorm, and the adjoint-derived observation-impact method allows for a deep-dive investigation into how various remote sensing observation platforms contribute to the intensity forecast of a high-impact east coast snowstorm.