Improving Stochastic Representations of Parameterized Physical Processes in Weather and Climate Models

A surprising finding in weather and climate modeling is that stochastically perturbing the parameterized diabatic physical tendencies of an atmospheric model improves its forecasts, its reliability, and its climatological variability. How these perturbations are specified is currently ad hoc and poorly motivated. To remedy this, we are developing a new rational methodology for diagnosing the stochastic components of model error in weather and climate models and subsequently improving them. To this end, we are using an hourly ensemble reanalysis and reforecast dataset developed with the NOAA Unified Forecasting System (UFS) and determining the space-time covariance statistics of its forecast error tendencies and its parameterized tendencies every hour for forecasts at leads from 1 to 12 hours over a full annual cycle. The space-time error covariances evolve in a manner originally suggested by Klinker and Sardeshmukh (1992), such that estimates of error evolution made using even 6 hourly tendencies are too far removed from the physical causes due to rapid adjustments in the atmosphere. One to two hour tendencies are essential. Using these, we observationally-constrain the space and time scales of the stochastic perturbations to be specified in the already beneficial stochastically perturbed parameterized tendency (SPPT) scheme. We find these space and time scales are meteorologically meaningful and are generally at least twice as large as what is currently used operationally. Results showing the impact of these larger scales in NOAA/UFS forecasts will be presented.