Moist adjoint-based forecast sensitivities for a heavy snowfall event over the Korean Peninsula on 4-5 March 2004

Adjoint sensitivity analyses are applied to a heavy snowfall event on the Korean Peninsula using the MM5 Adjoint Modeling System. To evaluate the effects of initial conditions on forecast error, adjoint integrations are performed using forecast error as a response function. Initial adjoint sensitivities are located in the middle to lower troposphere with horizontally isolated upshear-tilted structures in the upstream regions (i.e., southern Mongolia). In addition, due to the effect of moist physics in adjoint model integration, vertically striped structures in the lower troposphere are also detected in the southern sea of the Korean Peninsula. Both initial adjoint sensitivity structures match the individual singular vector structures. The new 36-h forecast using the initial condition perturbed with scaled adjoint sensitivities (i.e., key analysis error) shows much improvement in the intensity and location of the surface cyclone as well as the upper trough, and reduces 42.9% of the forecast error compared to the control forecast. In addition, the forecasts using the initial condition perturbed by the key analysis error deviate slightly from the observations than the forecasts using the control analysis in the first 12 h and are much closer to the observations than the control forecasts in the later 12 h. The linear and nonlinear evolutions of temperature perturbations over the large adjoint sensitivity regions show similar growth rates and spatial distributions. Additional experiments with and without specific moist physics for linear model integrations show that disagreement of moist physics in the nonlinear and linear model integrations degrades the linearity.