Goal: The primary goal of mesoscale data assimilation research is to develop and support state-of-the-art data assimilation systems for application in high-resolution mesoscale models. These data assimilation systems can be used for a variety of purposes including the assimilation of data from new observing systems, the optimal use of observations, and understanding the observational requirements for accurate precipitation forecasts and optimal strategies for obtaining targeted observations.

Development of MM5 3DVAR system

Dale Barker, Yong-Run Guo, Wei Huang and Qingnong Xiao have continued to add additional capabilities to the MM5 3DVAR system. In collaboration with Francois Vandenberghe (NCAR/RAP) and Shu-Hua Chen (post-doc visitor, now University of California, Davis), the direct assimilation of SSM/I brightness temperature has been coded and initial tests have been performed. In order to run 3DVAR in any particular area of the globe, regional background error statistics are required. A system is under development that will interpolate background errors (calculated via the NMC-method of averaged forecast differences) to a chosen domain. The system will then rescale them based on regional/resolution dependent observation/forecast difference values. Al Bourgeois has been working on portability and initial parallelization of the MM5 3DVAR system using the MPP framework of the WRF model. The code now successfully runs on a variety of platforms including the DEC Alpha, SGI, PC/Linux, IBM-SP, Fujitsu and NEC-SX5. An initial subset of 3DVAR has been parallelized and tested. Multiple/single processor results have been shown to be bit-reproducible.

Real-data applications of MM5 3DVAR system

The MM5 3DVAR system has been run in real-time at NCAR since July 2001 for the 135/45 km domains of the AOAWS MM5 model for the Taiwanese Civil Aviation Administration (CAA). More recently, the system has been ported to the CAA Fujitsu VPP5000 and run in real-time. Real-time applications provide an opportunity to test the accuracy and robustness of the 3DVAR system prior to operational implementation in 2002. A project between MMM and the Korean Meteorological Administration (KMA) was initiated in 2001. The goal is the eventual implementation of 3DVAR in the KMA's MM5-based operational regional data assimilation and prediction system (RDAPS). Collaboration among Yong-Run Guo, Dale Barker and Dong-Hyun Shin (KMA) has so far resulted in the porting of 3DVAR to the NEC platform, calculation of background error statistics for the Korean domains (via the NMC-method) and initial case-study assimilation of South Korea's high-density automatic weather station (AWS) surface observation network.

Continued development of MM5 4DVAR system

The deployment of radar and satellite remote sensing systems offers great potential to improve numerical simulations of the weather by improving the initial state. There is a significant obstacle to the use of radar and satellite data because the quantities that can be measured by these instruments are not directly usable by the models and tend to be irregularly distributed in time and space. Four-dimensional variational (4DVAR) data assimilation allows these observations to be assimilated directly into the forecast model. However, a principle disadvantage of 4DVAR is its large computational requirement, due to the iterative nature of method and its heavy use of CPU and memory. This limitation has all but prohibited its use at operational weather forecasting centers, and previous experiments have been limited to very small domains or relatively coarse spatial resolutions. John Michalakes has begun a project, in collaboration with scientists at AER Inc., to produce a complete 4DVAR system optimized to run on highly-scalable distributed-memory parallel computers. The current MM5v3.4 forecast model will be the non-linear model in the new 4DVAR system. It is already coded to run efficiently on distributed memory multiprocessor machines. Developing the tangent-linear and adjoint models is accomplished with help from the automated Tangent-linear and Adjoint Compiler (TAMC) and by hand checking the results with visual comparisons the MM5v1 versions. The tangent-linear and adjoint models are then modified to run on distributed-memory multi-processor machines, using the parallelization techniques employed with the basic MM5 forward model. The project, funded by the DoD High Performance Computing Modernization Office, targets completion in January of 2003. At this time the new 4DVAR system will be made available to the MM5 user community and may also be implemented operationally at the Air Force Weather Agency. MMM expects follow-on work with AFRL to focus on WRF 4DVAR.Yong-Run Guo and So-Young Ha (NCAR/ASP graduate student) have implemented several improvements to the current MM5 4DVAR system. These include: (1) compatibility with MM5 Version 3 for both input and output, (2) addition of a penalty term in the cost-function to control high-frequency oscillation, (3) portability of MM5 4DVAR system to Linux PC computing platform; and (4) compatibility of the new land-use category of the latest version of MM5 release. In addition, several bug-fixes have been released. These new improvements allow the existing MM5 4DVAR system to be compatible with the latest release of MM5, and also allow the MM5 4DVAR system to be operated on high-end linux PC systems. These improvements have been released to the MM5 user community.

Assessing the impact of lidar wind data

Dale Barker and Qinghong Zhang (NCAR/ASP postdoc) have begun to contribute to a NOAA-funded project to determine the potential benefits of a space-based wind-finding lidar for regional NWP. This work is a collaboration among FSL, ETL, NCEP and NCAR/MMM. Initial MMM work has been in the calibration of the 11-day trajectory of the MM5 forecast reference run (defined as truth) in the OSSE. Results indicate that this forecast is a reasonable source for the calculation of simulated observations for later assimilation.

Assimilation of land-surface data

Fei Chen (NCAR/RAP) and Kevin Manning implemented a high-resolution land data assimilation system for the Pennsylvania State University/National Center for Atmospheric Research Mesoscale Model, version 5 (MM5). This system is based on a land surface model currently used in MM5 but is run separately from the atmospheric model itself. The purpose of this "offline" implementation of a land surface model is to assimilate, over a significant period of time (several weeks or several months), observed quantities that drive soil moisture and temperature fields, especially observed fields of radiation and rainfall. The final fields of soil temperature and moisture, representing the assimilation of several weeks or months of data, may then be used as initial lower boundary conditions for the atmospheric model.

Data assimilation and forecasting on the convective scale

Juanzhen Sun and Andrew Crook have investigated the sensitivity of storm forecasts with respect to initial conditions obtained through the 4DVAR data assimilation technique. The goal of the sensitivity study is to investigate the features that must be retrieved in order to produce a good forecast and the data required to obtain such features. Both simulated and observed radar data have been used in this study. In the idealized study, a 4-hour simulation of a supercellular convective storm was performed using a sounding obtained during the CASES-97 experiment and a warm bubble initiation technique. This control simulation was used to generate single-Doppler radial velocity and reflectivity data at a frequency of five minutes, similar to WSR-88D data. These data are degraded with regard to coverage and quality and assimilated by the 4DVAR system to provide initial conditions for the subsequent forecast. The environmental wind and moisture were also varied to test the sensitivity of the storm forecast with respect to the environmental conditions. The sensitivity study was performed at the early-growth stage and the mature stage of the storm. The accuracy of the subsequent forecast was then evaluated by its correlation with the control simulation. There are three major findings in this study: 1) the forecast is very sensitive to low-level moisture at the early growth stage, but not as sensitive during the mature stage; 2) the radial velocity observations play a more important role than reflectivity in retrieving the low-level convergence, which is one of the key features in determining the forecast skill; and 3) the lack of low-level radar observations is more detrimental during the mature stage than during the growth phase. Figure 8 shows the rain-water correlation with respect to forecast time for five experiments during the growth phase.

Radar data assimilation experiments

The simulated data experiments have shown that a successful short-term forecast of a supercell can be performed from initial conditions retrieved from simulated radar observations. To determine if this success carries over to real Doppler observations Sun and Crook have performed a number of tests using radar observations of a severe tornadic thunderstorm observed during the CASES-97 experiment. Reflectivity and radial velocity from successive five-minute scans were assimilated into a cloud model using the 4DVAR adjoint method. A number of assimilation and forecast experiments have been performed with varying large-scale conditions (shear and CAPE), assimilation length and microphysical parameters (rainwater fallspeed and evaporation rate). Their initial results suggest that the storm forecast is very sensitive to the large scale conditions as well as the parameterized evaporation rate. The parameterized evaporation allows for the retrieval of a cold pool at low levels.

Influence of added observations on analysis and forecast errors

Rebecca Morss continued studying how adding observations to improve atmospheric analyses can influence errors in analyses and numerical forecasts. In collaboration with Kerry Emanuel (MIT), she used experiments with idealized data assimilation systems and forecast models to demonstrate why adding observations can degrade some analyses and forecasts, even with accurate observations and a perfect forecast model. She also identified several circumstances in which current data assimilation systems are more likely to use observational information to improve atmospheric analyses and forecasts.

Assimilation of GPS radio occultation sounding data

Ying-Hwa Kuo in collaboration with Tae-Kwon Wee (COSMIC postdoctoral fellow) has performed a set of observing system simulation experiments to assess the potential impact of GPS radio occultation soundings from a COSMIC-like constellation on the regional analysis and prediction over the Antarctic. They first performed a 30-km natural run over a 72-h period. The natural run was then used to generate potential GPS radio occultation soundings from the proposed COSMIC constellation, with realistic orbit parameters. This allowed the simulated soundings to have a realistic distribution in time and space. The simulated COSMIC soundings were then assimilated into a 120-km MM5 model. They showed that the COSMIC GPS radio occultation soundings could provide major improvement in the quality of regional meteorological analysis of the Antarctic. The improvement in the regional analysis would have a significant impact on the quality of regional weather prediction. Compared with earlier studies over the mid-latitudes, the impact over the Antarctic region is appreciably more significant. The relatively large impact of GPS radio occultation soundings is attributed to two major factors: (1) the Antarctic is a data sparse region of the world, and (2) the mass field (which is measured well by the GPS system) dominates the geostrophic adjustment process over high-latitudes.