Clouds and precipitation remain the most critical forecast elements in numerical weather prediction. The impact on societal activities related to business and transportation is broad. Increasingly these activities are becoming more time-critical as our transportation and business arteries become more congested. This mandates that weather be monitored at more frequent intervals and forecasts be issued with shorter lead times.

The primary issue for short term forecasts is model spin-up time. Forecasters need predictions where clouds and precipitation are available from the earliest time periods. At the Forecast Systems Laboratory we are seeking solutions for running forecast models in computationally limited environments, yet still offering initial conditions that have accurate representations of the water environment. Recent work with a high resolution cloud analysis and focus on the analysis of water in all phases have allowed us to consider merging the derived water environment with model initial conditions. The classically applied variational method continues to provide a means for complex analysis problems of this type. By running a simple, offline cloud model which retrieves liquid and ice partitioning and an estimate of vertical motion from the observed clouds, the variational scheme is designed to accept cloud vertical motion estimates and ice and water content as observations. The cloud vertical motion "observations" are fully coupled to the three dimensional mass and momentum fields using dynamical constraints which minimize the Eularian tendencies in the u- and v-components and ensure continuity is satisfied everywhere in the domain.

The scheme performs the analysis on the increment from an input model background. The variational scheme allows for model error (daily compiled) to be input explicitly at each gridpoint. When these grids are used to initialize a non-hydrostatic model, we find that there is no spin-up time and consequently excellent continuity between the initial condition and subsequent forecasts. Validation indicates that the initialization scheme provides better forecasts of state and cloud variables for about 6-9 hours in a limited domain, compared to runs where the model is initialized with the background only. Another advantage is that it need be run only once (no nudging is necessary) and is computationally efficient. This paper is complementary to others in the conference fully describing the moisture analysis and verification statistics for the model.