A Model Instability Issue in the NCEP Global Forecast System Version 16 and Potential Solutions

Abstract. The National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS) version 16 encountered a few model instability failures during the pre-operational real-time parallel runs. The model forecasts failed when an extremely small thickness depth appeared at the model’s lowest layer when strong tropical cyclones made landfall. A quick solution was to increase the value of minimum thickness depth, an arbitrary parameter introduced to prevent the occurrence of extremely thin model layers, thus numerical instability. This modification solved the issue of the model's numerical instability with a small impact on forecast skills. It was adopted in GFSv16 to help implement this version of the operational system as planned.

Further investigation showed that the extremely small thickness depth occurred after the advection of geopotential heights at the interfaces of model layers. In the FV3 dynamic core, the horizontal winds at interfaces for advection are calculated from the layer-mean values by solving a tridiagonal system of equations in the entire vertical column based on the Parabolic Spline Method (PSM) with high-order boundary conditions (BCs). We replaced the high-order BCs with zero-gradient BCs for the interface-wind reconstruction. The impact of the zero-gradient BCs was investigated by performing sensitivity experiments with GFSv16, idealized mountain ridge tests, and the Rapid Refresh Forecast System (RRFS). The results showed that zero-gradient BCs can fundamentally solve the instability and have little impact on the forecast performances and the numerical solution of idealized mountain tests. This option has been added to FV3 and will be utilized in the GFS (GFSv17/GEFSv13) and RRFS for operations in 2024.