• Advanced data assimilation systems for community use
• Optimal use of existing observations & the potential benefits of new observing systems
• Optimal strategies for obtaining targeted observations in data sparse regions

Development of the WRF 3DVAR System

The WRF data assimilation system is the focus for the Mesoscale Prediction Group's data assimilation efforts and has already been applied in both operational, research and data assimilation training applications. Barker, Yong-Run Guo, Wei Huang and Qingnong Xiao have continued to extend the capabilities of the WRF 3DVAR system. The capability to assimilate a number of satellite-derived retrievals (SATEM thicknesses, SSM/T1 temperatures and SSM/T2 relative humidities) has been created. In conjunction, a study of the quality of these retrieved products has been performed indicating caution must be applied in their use.

Al Bourgeois and Barker successfully merged the WRF 3DVAR system with the WRF software framework in order for 3DVAR to run efficiently on distributed memory computing platforms. An example of recent progress in scaling the 3DVAR system to run on multiple nodes of NCAR's IBM-SP "blackforest" machine is shown in Fig. 12. The particular case studied is a 3DVAR run valid at 12Z on 25 January 2002 in AFWA's 45km South-West Asian "T4" 140x150x41 domain. Wall-clock time is reduced from ~1370s (1 processor) to 89s (64 processors). Additional tuning of background errors results in a reduction to 58s on 64 processors. Although scalabilty is not perfect, a run-time of under 1 minute is well within the time window of current operational and research applications.

Figure 12. Scalability test for 3DVAR run using AFWA's 140x150x41 45km T4 domain with the 25 January 2002 case study. Times shown are for runs on NCAR's IBM-SP machine blackforest with Winterhawk II nodes.

Barker has compared numerous methods for the computation of observation and forecast error covariances. These include the default "NMC-method" of forecast error covariances approximated via averaged forecast differences, error estimation via fit of forecasts to observations, and adaptive parameter estimation methods which utilize 3DVAR analyses assimilating randomly perturbed observations. The latter method has been shown to be successful in automatically tuning observation/forecast error statistics, in response to changes in observation network and/or forecast models.

The 3DVAR component of WRF is scheduled for official release to the research community at the end of 2002. Support for this new community data assimilation system will be key to its success. Even before its release, the code had been used by collaborators to assess the impact of new observation types (e.g., SSM/I and ground-based GPS). This type of work is expected to grow as the number of users increases.

Real-data applications of WRF 3DVAR

The WRF 3DVAR system has been successfully implemented in two operational environments in 2002. First, after months of testing and tuning, 3DVAR was officially delivered to the Taiwanese Civil Aviation Authority (CAA) in May 2002. The system runs as part of the MM5-based Advanced Operational Aviation Weather System (AOAWS). Secondly, on 12 September 2002, WRF 3DVAR was made operational in numerous "theaters" world-wide at the United States Air Force Weather Agency (AFWA). These 3DVAR implementations represent the first use of a component of WRF in an operational environment.

Figure 13 illustrates one month's Mesoscale Model 5 (MM5) forecast geopotential height verification against radiosondes for AFWA's "T3" (Europe) 45-km domain. The improvement using 3DVAR (triangles) versus the previously operational optimum interpolation (OI) scheme (squares) is clearly seen at both 12-h (green) and 24-h (blue) forecast ranges. Not only is 3DVAR producing superior forecasts, but the 3DVAR implementation at AFWA is effectively computationally cost-free. This particularly satisfying achievement is due to 3DVAR's ability to "cycle" (i.e., use previous MM5 forecasts as a starting point for the analysis). The 3DVAR cycling minimizes spin-up problems (e.g., insufficient cloud) seen previously in the early hours of OI/MM5 forecasts. The use of 3DVAR permits the elimination of an additional 6-h MM5 forecast (previously required to reduce spin-up via a "nudging" procedure), saving more CPU than is required by 3DVAR itself.

Figure 13. MM5 forecast height verification against radiosondes for the AFWA's "T3" (Europe) 45-km domain during the period 4 June-10 July 2002. Initial conditions supplied by 3DVAR (triangles) and the previous operational optimum intepolation (OI) scheme (squares) at hour 0 (red), hour 12 (green), and hour 24 (blue) forecast ranges are shown.

Operational Application of MM5 3DVAR at KMA

Related website: http://210.107.255.13

Starting in April 2001, the MMM Division started to collaborate with the Korea Meteorological Administration (KMA) on the development of the MM5 3DVAR system, and its application in operational regional numerical weather prediction. Through collaboration with KMA staff, Barker and Guo implemented a version of the MM5 3DVAR into the KMA operational system. Guo implemented the capability to assimilate observations from Automatic Weather Stations (AWS) over the Korea peninsula. The semi-operational runs have been conducted in Korea Meteorological Administration (KMA) since June 2002. A web site: http://210.107.255.13 was setup to display the real-time results from the MM5 3DVAR cycling run in KMA. This allows a side-by-side comparison with KMA's operational three-dimensional OI scheme (3DVOI) on a routine basis. The results have shown that MM5 3DVAR performs, in general, better than the operational 3DOI scheme. In particular, the use of MM5 3DVAR improves the operational model's skill in quantitative precipitation forecast. Moreover, the assimilation of AWS data is shown to have a positive impact on the precipitation forecast over South Korea.

Development of Radar Data Assimilation with MM5 3DVAR system

Beginning in June 2002, MMM and KMA started to collaborate on the assimilation of Doppler radar data into the MM5 3DVAR system. The goal is to assimilate both radial velocity and radar reflectivity data, from Doppler radars around the Korean Peninsula, for mesoscale model (10-km grid size) initialization, and to assess the impact of Doppler radar data assimilation on short-range precipitation prediction. To enhance the capability of the MM5 3DVAR system for radar radial velocity assimilation, Xiao and Barker improved the MM5 3DVAR system by including vertical velocity increments in the analysis. Wen-Chau Lee (NCAR/ATD) performed radar data quality control for a squall line case that took place over the Korean Peninsula on 10 June. Juanzhen Sun implemented a radar data pre-processor for MM5 3DVAR. Xiao, Sun, and Guo developed the radial velocity observation operator for the system. Numerical experiments are now being carried out to test its capability and to assess the impact of assimilating radial velocity using a squall line that occurred on 10 June 2002. Different strategies for radar data assimilation, including cycling at every 3 hours and every hour, will be tested.

Development and Testing of a Parallelized Version of MM5 4DVAR

A major hurdle for using 4DVAR in operational mesoscale data assimilation is the computational cost. The existing MM5 4DVAR system was developed based on MM5 version 1.0. That code could only be run on a single CPU and could not take advantage of either the shared- or distributed-memory parallel computers. As a result, the MM5V1 4DVAR system took prohibitively long periods of wall clock time for data assimilation or adjoint sensitivity calculation. Under the sponsorship of the High Performance Computing Office of the U.S. Department of Defense, John Michalakes, in collaboration with researchers at Atmospheric Environment Research (AER) Corporation and the Air Force Research Laboratory (AFRL), developed a parallelized version of 4DVAR based on the latest version of MM5.

Guo assisted with the testing of the new MM5V3 4DVAR system. Initial testing certified correctness of the tangent linear and adjoint model, once a few bugs were fixed. Intensive tests were subsequently performed to assess the efficiency of the code. With the INTERNAL-IO technique implemented by Michalakes, the speed-up factors reached 65.99 on IBM SP P3 and 196.03 on Compaq ES-45 for a large case (SESAME: 136x152x10, grid size; 15-km, and 6-h assimilation window). This shows the new parallelized code has the potential for operational mesoscale data assimilation. Guo also compared the results from this CWO-5 Beta code to the MM5V1 4DVAR system. The increments of the meteorological fields are consistent with each other as illustrated in Fig. 14.

Figure 14. Comparison of analysis increment produced by MM5 V1 4DVAR with MM5 V3 (CWO-5) for the SESAME-I case.

Assessing the impact of simulated space-based lidar observations on regional NWP

Barker and Qing-Hong Zhang (NCAR/ASP) continued to collaborate with NOAA/Forecast Systems Laboratory on the testing and verification of forecasts run as part of an observation system simulation experiment (OSSE), to assess the potential impact of a space-based lidar instrument. Having verified the suitability of an MM5 nature run, they began to verify RUC forecasts using simulated conventional data.

Data assimilation experiments with Doppler radar data (4DVAR and EnKF)

Crook, in collaboration with Dowell, pursued a data assimilation comparison study using Doppler radar observations of the Arcadia, Oklahoma supercell. This was a tornadic supercell that occurred on 17 May 1981 and was well sampled by two Doppler radars. Crook applied a 4Dvar technique to assimilate single Doppler radar observations into the cloud-scale numerical model developed by Juanzhen Sun (NCAR/MMM/RAP). Dowell used a second technique, the Ensemble Kalman filter (EnKF), to assimilate the same radar observations. It was found that both techniques are capable of assimilating single Doppler radar observations into a numerical model and producing reasonable convective scale structures (in terms of cross-beam velocity, thermodynamics and microphysics). Preliminary results suggest that the 4DVAR results verify better (against the Doppler measurements from the second radar) in the early part of the assimilation cycle but that EnKF does better later in the cycle. While the assimilations are clearly far from perfect at this early stage, they find evidence that the EnKF is extracting useful information from the observations: assimilating observations from one radar results in a steady improvement, over about 30 minutes, of the fit of the analysis to the observations from the other radar. Finally, short-term (one to two hour) forecasts have also been performed from the retrieved initial conditions. These forecasts have been successful in terms of producing a smooth startup as well as predicting the observed storm motion, but less successful in predicting the storm evolution past one hour.

Comparison of 4DVAR and EnKF techniques for convective-scale data assimilation

Sun, Caya, and Snyder conducted a study to compare the performance of the 4DVAR and EnKF techniques using simulated data of a supercell storm. 4DVAR and EnKF are the two analysis techniques capable of accounting for the evolution of the atmospheric flow represented by a numerical model. The 4DVAR technique was examined for the convective-scale in the past and showed promising results. The EnKF is a relatively new technique with some attractive features, such as it takes less effort to implement compared with the 4DVAR technique. The goal of Sun and her colleagues' study is to determine whether the EnKF technique has similar performance and robustness to the 4DVAR technique. Using a simple cloud model and radar "observations" extracted from the simulation of a supercell storm, identical-twin experiments were conducted by both the 4-D variational Doppler radar analysis system (VDRAS) and a recently developed EnKF analysis system. The preliminary results suggest that both of the techniques are able to retrieve the unobserved wind, thermodynamical, and microphysical variables with good accuracy when the systems were carefully tuned.

Assimilation of Surface-layer Observations Using EnKF Approach

Surface-layer in-situ observations are a rich, accurate, and often dense data source; however, they are generally under-utilized in current operational data assimilation (DA) systems because complex interactions between the surface and the atmosphere aloft are difficult to handle. Using these observations to improve the initial planetary boundary layer (PBL) state in a mesoscale NWP model could lead to better short-range forecasts of convective outbreaks, slope flows, and frontal propagation. Research by Hacker, Snyder, and Anderson uses an ensemble Kalman filter (EnKF) to assimilate idealized, in-situ, mesonet observations into a PBL model. Because its response is expected to be similar to the response in a mesoscale model, the model chosen is a PBL parameterization scheme commonly employed in mesoscale models. It is considered perfect for initial experimentation. The EnKF approach is promising because an ensemble of mesoscale forecasts can develop anisotropic background error covariances. Hacker et al. demonstrated proof-of-concept by running the PBL model in 1D and assimilating surface-layer temperatures. The resulting state of the PBL was more accurate than a state estimated with an idealized DA system as shown in Movie 2. An error term, representing both model error and forecast error that may be intrinsic to the real atmosphere, was also introduced with positive results.

	To view the movie, place mouse over image. Alternately, for slower connections, you may use the links below to download the movie. PBL movie (animated GIF) PBL movie (AVI format) PBL movie (quicktime format)
Movie 2. Potential temperature profiles of the PBL, comparing no assimilation, ensemble Kalman Filter (EnKF) assimilation, and idealized assimilation (FDDA) of 2 m temperatures. A single case is shown. No assimilation occurs before time ta and the assimilation remains active until the end of the simulation. The thick black line is the ensemble mean, the thick blue line is the control ("truth"), and the red lines show 10% of the ensemble members.

Initialization and forecasting of a hailstorm using radar observations from STEPS

Forecasting of individual convective systems presents great challenges to numerical weather prediction. In order to examine the ability of qualitative prediction of precipitation in warm season convective systems, Sun and Miller initialized a warm rain cloud-scale model using single-Doppler observations of a hailstorm from a WSR-88D radar and performed short-term forecasting experiments. The model was initialized by assimilating the radar observations for a period of 20 minutes, using the 4DVAR technique. A series of data assimilation and forecasting experiments were conducted to examine the sensitivity of the forecast, with respect to environmental, model, and data uncertainties. Sun and Miller's results indicate that the gross structure, location, and propagation speed of the predicted storm have good agreement with the observations for up to two hours. The sensitivity experiments suggest that the forecast accuracy highly depends on the environmental low-level moisture and the microphysical parameterization.

Assimilation of Slant-path Water Vapor Measurement with Ground-based GPS Network

With the recent advance in Global Positioning System (GPS) sensing technology, slant wet delay along each ray path can be measured to within a few millimeters accuracy. Potentially, this can provide high-temporal resolution water vapor measurements at low cost. So-Young Ha (NCAR/ASP), Kuo, and Guo performed a series of observing system simulation experiments to assess the impact of slant wet delay on the short-range prediction of a squall line. Specifically, they assimilated slant wet delay data from a hypothetical network of ground-based GPS receivers using the MM5 4DVAR system. They showed that the assimilation of slant wet delay results in significant changes in moisture, temperature, and wind fields, within the boundary layer. These changes lead to a stronger surface cold front and stronger convective instability ahead of the front. Consequently, the assimilation of slant wet delay produces a considerably improved 6-h forecast of a squall line, in terms of rainfall prediction and mesoscale frontal structure.

In order to assess the additional value of slant wet delay assimilation (as compared with zenith wet delay which is equivalent to precipitable water measurements), Ha, Kuo, and Guo performed a parallel experiment in which precipitable water is assimilated. They demonstrated that the assimilation of slant wet delay is superior in recovering water vapor information between receiver sites, and in short-range precipitation forecast, both in terms of rainfall distribution and intensity. The assimilation of slant wet delay more accurately retrieves the temperature and moisture structure in the convectively unstable region.

Impact of Digital Filter as a Weak Constraint in MM5 4DVAR

The assimilation of heterogeneous observations, using the 4DVAR system, is known to generate high-frequency intertia-gravity waves. These high-frequency waves are caused by a variety of sources: 1) insertion of data causing dynamic imbalance; 2) observational errors; and, 3) incompatibility between model and observations. These high-frequency oscillations are not meteorologically significant, and can reduce the effectiveness of 4DVAR. Since inertia-gravity waves are acceptable solutions to a 4DVAR system, the 4DVAR minimization may try to fit the observations to these high-frequency waves. As a result, the overall effectiveness of 4DVAR is reduced. Tae-Kwon Wee (UCAR/COSMIC) and Kuo have introduced a digital filter into the MM5 4DVAR system, as a weak constraint, in order to control high-frequency oscillations which negatively impact assimilation performance. To assess the impact of digital filters, and to understand how the digital-filter 4DVAR functions, a series of Observing System Simulation Experiments (OSSE), with the assimilation of Global Positioning System (GPS) refractivity soundings for a cyclogenesis case over the Antarctic region, were conducted. They showed that the weakly constrained 4DVAR with digital filter not only reduces dynamic imbalance, but also significantly improves the qualities of analysis and forecast (Fig. 15). Without projecting its solution onto the high frequency waves that diminish rapidly with forecast time, the constrained 4DVAR is able to yield additional improvement in model initial condition in the larger-scale range. Hence, it utilizes the available observations more effectively compared to the unconstrained 4DVAR. Wee and Kuo also showed that the filtering of wind field is found to be the most effective in suppressing high-frequency oscillations.

Figure 15. Vertical velocity fields (cm/s) on the second layer from the model top at the end of assimilation period in No-Digital-Filter (left panel) and in Digital-Filter (right panel) experiments.