I'm not sure what you mean about "compensating" for the lack of 0/12Z soundings. They have any soundings from 06/18Z, plus satellite soundings, aircraft soundings, etc. They also use parts of previous runs, depending on the model.
Take a look at
http://www.meted.ucar.edu/topics_nwp.php / Understanding Assimilation Systems: How Models Create Their Initial Conditions for more detail which was just published (along with a lot of other model training on their homepage) last week.
Description:
Understanding Assimilation Systems: How Models Create Their Initial Conditions, is part of the NWP Distance Learning Course: "Effective Use of NWP in the Forecast Process." This module explains the data assimilation process, including the role of the model itself as well as the observations. It provides learners an appreciation for how models use data as a function of model resolution and data type, how data influence the analysis, the limitations of data assimilation systems, the importance of initial conditions on the quality of NWP guidance, as well as the challenges of assessing the quality of NWP guidance based on the initial conditions. The differences between 3d-var with isotropic background covariances, anisotropic background covariances, 4d-var, and ensemble Kalman filter are conceptually illustrated.
Objectives:
After completing this module, students should be able to:
1) State the 2 primary sources (types) of information used in making a model
analysis
2) List 3 advantages of using a model short-range forecast as a first guess for the analysis
3) State the fundamental assumption of data assimilation in NWP models and
describe something which can happen if this assumption is violated. The answer to this question is the same regardless of the assimilation method (e.g., 3d-var, 4d-var, ensemble Kalman filter, etc.)
4) Describe what can happen to the analysis and forecast when a model can resolve and create features smaller than the network of observations assimilated can resolve in the vicinity of the feature
5) State one reason why a model might produce a better forecast if its analysis
does not fit a perfect observation too closely
6) Given observations of a weather feature showing a first guess error in
placement of the feature, describe how the feature will differ in a 3d-var analysis, 4d-var analysis, and ensemble Kalman filter analysis
7) List 4 scenarios that can produce a poor analysis leading to a poor forecast
and describe the cause or nature of the analysis problem
Estimated time to complete: 2.00 - 3.00 h
