Smolarkiewicz, Vanda Grubisic [NCAR's Advanced Study Program (ASP)], and Margolin continued the three-year study funded by the DOE Computer Hardware, Advanced Mathematics and Model Physics (CHAMMP) program. This is a study of nonoscillatory forward-in-time (NFT) methods suitable for modeling the global dynamics of the atmosphere and oceans. Last year this team extended the semi-Lagrangian variant of the NFT fluid model on a mountainous sphere and dispensed with the traditional geophysical simplifications of hydrostaticity, gentle terrain slopes, and weak rotation. This resulted in perhaps the first fully nonhydrostatic NFT model for a rotating, mountainous globe. Because NFT methods are inherently two-time-level, the accurate integration of forces leads to a complex nonself-adjoint 3D elliptic partial differential equation (PDE) for pressure. For atmospheres of depth comparable to the radius of the sphere, such equations can be solved easily using appropriate iterative methods. For thin-shell fluid problems typical of Earth meteorology, however, the resulting elliptic PDE is extremely stiff. One remedy is to use the hydrostatic solution for pressure as a guess for the iterative nonhydrostatic pressure solver. This leads naturally to a code design that facilitates the optional implementation of hydrostatic or nonhydrostatic models, leaving the final choice to the user.
William Anderson and Smolarkiewicz continued their work on high-performance computing strategies for atmospheric models. Last year they implemented a three-dimensional nonhydrostatic, time-dependent Navier-Stokes solver using two different parallelization methods, High Performance Fortran (HPF) and message passing (MSG). The solver has two distinct (optional) fluid algorithms, a Eulerian and a semi-Lagrangian, with different numerical properties and communication patterns. For each algorithm, they compared the performance of the two parallization methods on a Cray T3D. For the same time step of integration, the Eulerian algorithm runs 1.3 and 1.7 times faster than the semi-Lagrangian scheme for, respectively the HPF and MSG versions. The MSG code runs 2.5 and 1.8 times faster than the HPF code for, respectively the Eulerian and semi-Lagrangian versions. Since the semi-Lagrangian model admits large time steps, its overall performance is better than that of the Eulerian model (at least for the class of stratified flows past complex terrain addressed in their study), despite its inherent irregular communication patterns.
The adaptive grid model COMMAS (COllaborative Model for Multiscale Atmospheric Simulation), using the generalized adaptive grid interface constructed by William Skamarock and Ming Xue [Center for the Analysis and Prediction of Storms (CAPS)], a cloud model constructed by Louis Wicker (Texas A&M University) and Robert Wilhelmson (National Center for Supercomputing Applications, University of Illinois), and a sophisticated ice-microphysics parameterization were used for simulations of midlatitude mesoscale convective systems, supercell storms, coastal currents, and frontogenesis associated with unstable baroclinic waves. A terrain version of COMMAS is being developed, in cooperation with Dale Durran (University of Washington), that uses an advective, governing-equation formulation that has proved to be the most robust system for use in cloud-, meso- and synoptic-scale simulations.
Skamarock, Joseph Klemp, and Smolarkiewicz continued to examine alternative methods for integrating various forms of the nonhydrostatic equations. Semi-implicit integration methods, incorporating both semi-Lagrangian and Eulerian approaches, were examined for accuracy, robustness, and efficiency for cloud- and meso-scale applications. In particular, efficient Helmholtz and Poisson equation solvers were developed for the nonconstant coefficient problem that includes cross-derivative terms. These solvers are based on the conjugate residual (CR) algorithms developed by Smolarkiewicz, and preconditioners were identified that accelerate the algorithms such that the full models are as efficient as existing time-split models or models based on direct solvers for the constant coefficient problem.
Wicker and Skamarock began to examine a new class for split-explicit integration methods of integrating the elastic nonhydrostatic equations that offer some advantages over the traditional leapfrog-based split-explicit approaches, including two-time-level simplicity and a decreased need for explicit computational filtering.
Mesoscale Prediction Group (MPG) and Mesoscale Dynamics Group (MDG) scientists began a joint effort to develop a new mesoscale forecast model that will be well suited for both idealized modeling research and numerical weather prediction (NWP) applications across a broad range of scales. This model would replace the current NCAR/Pennsylvania State University (PSU) Mesoscale Model (MM5) maintained by MPG as well as the cloud models used by MDG. To develop closer ties with the operational forecast models, MMM is exploring the possibility of pursuing this effort as a cooperative project with the National Oceanic and Atmospheric Administration's (NOAA) National Centers for Environmental Prediction (NCEP) and Forecast Systems Laboratory (FSL) to develop a joint model that would be used both for research in the university community and for operational forecasting. A team of NCEP/FSL/MMM scientists will explore this opportunity in the coming year, and seek to define an optimal numerical framework for the new model, currently referred to as the Weather Research and Forecast (WRF) model, that will meet their respective requirements.
As a preliminary step toward determining the numerical approaches for the new model, Klemp and Skamarock began an investigation of the strengths and weaknesses of alternative techniques for treating orography using simple idealized flow conditions. They found that in two-dimensional simulations of mountain waves using a terrain-following transformed z (or sigma) coordinate, the waves are well resolved provided the vertical grid interval is small compared to the vertical wavelength of the mountain waves. With a step-mountain representation of the terrain, good resolution of the waves requires the vertical grid interval in the vicinity of the mountain to be small compared to the height of the terrain, which is a much more restrictive requirement. Without high vertical resolution along the terrain, the step-mountain approach may generate significant disturbances at the corners of each step that propagate well into the troposphere. Experiments are also being conducted with the terrain-following coordinates for a resting atmosphere adjacent to steep terrain to assess the significance of pressure gradient errors in that coordinate framework.
The accuracy of forecasts for the Western United States is often limited by the sparsity and quality of upstream data over the Pacific Ocean. An intriguing potential remedy for this problem is an adaptive observing strategy, in which movable observing platforms (such as manned or unmanned aircraft) provide additional observations in those specific regions where, depending on the character of the flow at a given time, analysis errors are likely to be large and to grow rapidly. Following a workshop convened by Christopher Snyder at NCAR in May 1995, international interest in this problem has blossomed.
Kerry Emanuel [Massachusetts Institute of Technology (MIT)], and Snyder will test adaptive strategies during the Fronts and Atlantic Storm Track Experiment (FASTEX) in January and February 1997. Drop soundings will be performed by a Lear jet flying from St. Johns, Newfoundland; this aircraft will be leased from Flight International, Inc. and funded by the National Science Foundation (NSF)/NCAR deployment pool. (Emanuel and Snyder originally proposed to bring the NCAR WB57F to FASTEX, but mechanical difficulties currently limit the use of this airplane for research.) During the experiment, the adaptive strategies will be based on bred vectors produced by the NCEP or on singular vectors calculated at the European Center for Medium-Range Weather Forecasts (ECMWF) and at the Naval Research Lab (NRL). A significant portion of other FASTEX resources also will be committed to testing adaptive observing strategies. National Oceanic and Atmospheric Administration (NOAA), Meteo-France, and NRL have all contributed to funding the NOAA Gulfstream-IV airplane for this purpose.
Rebecca Morss (MIT) and Snyder began testing adaptive strategies through observing system simulation experiments (OSSEs) with a quasi-geostrophic model. During the past year, they developed and tested a three-dimensional variational data-assimilation scheme for this quasi-geostrophic model. Using this scheme, Morss showed that an adaptive strategy based on bred vectors provided a greater increase in forecast skill than the same number of additional observations with randomly chosen locations.
In order to assess the impact of actual Global Positioning System/Meteorology (GPS/MET) refractivity data on numerical weather prediction, Xiaolei Zou, Wei Huang, and Ying-Hwa Kuo performed a series of four-dimensional variational data assimilation (4DVAR) experiments during a 12-hour period on 14 October 1995. This is within a period when "anti-spoofing" was off, and GPS/MET data of higher quality were obtained. A total of 10 GPS/MET occultation soundings was included within the model domain, which covers the contiguous 48 states, Canada, and Mexico. Analysis of the 4DVAR results indicate that the MM5 4DVAR system can effectively assimilate the actual GPS/MET refractivity data into the model. The 4DVAR is an effective procedure to retrieve temperature and moisture profiles from the GPS/MET refractivity data. Preliminary results also indicate that the assimilation of GPS/MET refractivity data improves the prediction of an extratropical cyclone over the west coast of the U.S.
It is well known that the GPS/MET raw refraction angle data provide better simulation conditions than the refractivity profiles, because the assimilation system can more accurately simulate the GPS data which implicitly contains spherical symmetry assumptions. In close collaboration with Mikhail Gorbunov and Sergey Sokolovskiy from Russia, and Simon Rosenfeld (NCEP), Zou, and Francois Van den Berghe (visitor from Cabinet d'Etude Technique Industrielle et d'Innovation Scientifique, France) began developing a procedure to directly assimilate the raw GPS refraction angle data using the NCEP global spectral model. Over the past year, Zou and Van den Berghe completed the following tasks: (1) modification of the NCEP model and its adjoint model to run the current operational NCEP data; (2) testing of the adjoint of various physical processes and their linkage to the dry 4DVAR system developed several years ago; and (3) development of observation operator software needed to carry out direct assimilation of the GPS refraction angle along with a high accuracy interpolation algorithm. In the coming year, tests will be performed using the actual GPS/MET refraction angle data collected in October 1995 with the NCEP global spectral model.
Working closely with Joseph Sela and John Derber (both of NCEP), Zou continued to develop the 4DVAR system based on NCEP's global spectral model. A full-physics adjoint model of the NCEP global spectral model was established over the past year. In the coming year, we will make use of the NCEP global and its 4DVAR system for GPS/MET data assimilation research as well as research in support of the U.S. Weather Research Program (USWRP) and North American Observing System (NAOS) programs. This effort will also benefit the collaboration between NCEP, FSL, and MMM on the development of the WRF model previously mentioned.
Zou, Huang, and Yong-Run Guo continued to develop the MM5 full-physics adjoint and the MM5 4DVAR system. The MM5 adjoint model includes the following physics options: large-scale condensation; microphysical parameterization for cloud water; rain water and ice; subgrid-scale convective parameterization of Grell and Kuo schemes; atmospheric radiation; surface flux parameterization; Bulk PBL scheme; and Blackadar PBL scheme. These physical parameterizations, together with the nonhydrostatic dynamics of MM5, allow the MM5 adjoint model to be a highly versatile tool for mesoscale dynamics study and data assimilation research. Over the past year, the MM5 4DVAR system was used successfully for GPS/MET refractivity assimilation, rainfall data assimilation, and precipitable water data assimilation. We also put significant effort into making this MM5 adjoint model and 4DVAR system more user friendly. We plan to officially release the MM5 adjoint and 4DVAR system to the user community in the summer of 1997.
In collaboration with George Modica and Alan Lepton (both from the Air Force Phillips Laboratory), Zou and Qinglong Xiao (visitor from Nanjing University, China) developed tangent-linear and adjoints of the radiative transfer model for Special Sensor Microwave Imager (SSM/I), Special Sensor Microwave Temperature (SSM/T), and SSM/T-2 observations. One-dimensional variational retrieval tests with simulated SSM/T and SSM/T-2 data were carried out, and the forward radiative transfer model will be installed in the MM5 adjoint system. The ability to perform satellite data assimilation using the MM5 4DVAR system opens many opportunities for future research.
Zou and Kuo continued their work on rainfall data assimilation. They performed additional experiments to assess the impact of precipitation parameterization on rainfall data assimilation. They found that the results of rainfall data assimilation were not very sensitive to cumulus parameterization schemes. The 4DVAR rainfall assimilation experiments using the Grell and Kuo schemes produced very similar adjustments to the original analysis. Moreover, rainfall data assimilation produced considerable impact even if different schemes were used in the assimilation and forecast periods. Further experiments also showed that inclusion of more sophisticated cloud microphysics for the resolvable scale precipitation resulted in further improvements. These results suggest that the differences in model prediction caused by different cumulus parameterization schemes are smaller than the analysis errors in the model initial conditions. They also demonstrate that improved quantitative precipitation forecasts are possible through the assimilation of rainfall observations, along with other conventional data.
Christopher Davis, Kevin Manning, and Zou performed exploratory simulations to test the viability of assimilating single-Doppler radar data for improving the short-term precipitation forecast with MM5. The short-term precipitation forecast (0-6 hours) is hampered by model spinup problems, particularly resulting from a lack of dynamical consistency between vertical motion, moisture and microphysical constituents. A particularly intense snowband, observed during Winter Icing and Storms Projects, 1994 (WISP94) was studied. To test feasibility of adjoint initialization, the Observing Systems Simulation Experiment (OSSE) approach was used, in which the "observations" were generated by a 20-km resolution forecast. This simulation produced a snowband similar to the observations. From these data was computed a radial wind relative to a fixed point at the ground (Dodge City, Kansas). Information was only used from saturated grid boxes; no information was available where the atmosphere was unsaturated. A 60-km resolution forecast, initialized 12 hours earlier than the simulation which produced the observations, provided a background field, devoid of a snowband. The simulations performed sought to determine how much of the observed precipitation could be predicted in a simulation whose initial conditions consisted of the coarse background enhanced only by knowledge of the radial wind and the regions of saturation over a finite area. The prediction of heavy precipitation in the 0-4 hour forecast was greatly enhanced over that appearing in the background. For areas with precipitation rates exceeding 1 mm per 20 minutes, equitable threat scores approached 0.5 in the first hour, and dropped off to about 0.1, the level of the background, by four hours. The time scale over which the precipitation forecast was improved appeared to be driven mainly by errors in boundary conditions. To help alleviate the boundary condition problem, simulations were performed which mimicked a mosaic of radars over the Western High Plains. These results were generally consistent with the previous studies, although the region experiencing an improved precipitation forecast was greater.
David Parsons (ATD), Jimy Dudhia, and Guo continued their collaborative data assimilation research using data collected in a ten-day field program in June 1993, with an objective to provide a dynamically consistent four-dimensional analysis over the Atmospheric Radiation Measurement (ARM)'s Southern Great Plains (SGP) study site, which can be used for various types of diagnostic studies. Over the past year, they produced a set of four-dimensional analysis based on the nudging technique, and successfully validated the resulting data set against independent observations. This data set is currently employed to validate cloud and radiation parameterizations used in general circulation models for climate studies, in collaboration with Jonathan Petch and James Hack [both of NCAR's Climate and Global Dynamics Division (CGD)].
Davis and Manning continued to examine the performance of the MM5 model in its operational forecasting during WISP94. Recent work emphasized the performance of the PBL scheme. Using two months of rawinsonde data over the Central and Western U.S., it was shown that MM5 failed to capture the strength of the nocturnal inversion, and showed a cold and moist bias in the PBL during the day. The result is a diurnal cycle of smaller amplitude than observed. Collaboration with Dudhia and James Bresch (visitor, University of Washington) confirmed the sensitivity of the PBL treatment to the specification of soil moisture.
Bresch continued to develop an experimental real-time MM5 forecast system that utilizes the initial conditions and forecast lateral boundary conditions from the NCEP Eta model. The forecasts, which run once or twice daily on the Division's multi-processor machines (SGI Power Challenge, Cray J-94), are for 36 hours, and are centered on Colorado at 27-km grid size with options of a 9-km nest or an alternative 27-km run that is either relocated or has different physics. The real-time forecast system was used to provide forecasts for the Stratosphere-Troposphere Experiment: Radiation, Aerosols, and Ozone (STERAO) field program over northeastern Colorado in the summer of 1996. The daily MM5 forecasts are available on the MMM Web page, including a variety of severe storm parameters that are being evaluated for their predictive usefulness (in collaboration with Erik Rasmussen, NSSL).
Simon Low-Nam, David Gill, Alexis Lau (visitor, Hong-Kong University of Science and Technology) Wei Wang, Xin An Chen (visitor, University of Hawaii), and Daniel Hansen completed an operational mesoscale forecasting system based on MM5, which operates on a high-end workstation platform. This operational MM5 real-time forecasting system passed the Factory Acceptance Test set up by the Royal Observatory in September 1996. The system includes a six-hour update cycle, which uses its previous forecast as the first guess to the analysis for the next cycle. The first guess fields are enhanced by the incoming data collected in real-time from the Global Telecommunication System. The data are quality controlled and objectively analyzed to the model domain after a two-hour data cutoff. A six-hour pre-forecast dynamic initialization is then performed, by nudging the model predicted variables towards the new analysis, to produce a final analysis from which an 18-hour forecast is made. The lateral boundary condition for this regional system is provided by the NCEP AVN (Aviation) model available from the World Area Forecast Centers. The MM5 forecast is used as an input to the Operational Windshear Warning System (OWWS) being developed for topographically induced windshear and turbulence prediction at Chek Lap Kok airport, Hong Kong.
Observational studies have shown that the strong southerly and southeasterly flow crossing the Lantau Island when a typhoon approaches Hong Kong are conducive to strong topographically induced windshear and turbulence. Typhoons also produce floods, landslides and gusty winds that pose a great threat to properties and lives in Hong Kong. Therefore, an accurate prediction of the intensity and track of a typhoon approaching Hong Kong is highly important to the development of the OWWS system. The initialization of a typhoon in a mesoscale model is not an easy problem. First, operational global analysis cannot correctly describe the vortex circulation associated with the typhoon. Second, there are little or no observations in the vicinity of a typhoon to correct errors in the analysis. The problem is further complicated with the use of mesoscale update cycles. For instance, significant errors may exist in terms of the position and intensity of a typhoon in the previous forecast cycle. Such errors cannot be removed due to the lack of observations. As part of the development of the MM5 real-time forecasting system, Low-Nam, Chen, and Kuo developed a procedure to remove erroneous vortex circulation in the previous update cycle and implant a new vortex with the correct intensity and position in the final analysis. The vortex implantation procedure is shown to produce dynamically balanced fields without altering the large-scale circulations. This procedure is found to significantly improve the skill of MM5 in typhoon track and intensity prediction.
Following last year's demonstrations of a numerical weather prediction system operating via satellite between remote supercomputers, Jordan Powers (ASP) modified the MM5-based coupled modeling system to run on supercomputers located at NCAR's two Boulder sites, the Foothills and Mesa laboratories. This was done as part of the Advanced Research Projects Agency (ARPA)/COOP-3D project, which involved the operation of a coupled mesoscale modeling system employing a distributed computing design. The intra-NCAR connectivity was provided by an Asynchronous Transfer Mode (ATM) network, and the supercomputers were Cray J-90's in MMM and SCD. Different MM5 forecast domains ran on the separate machines, and the simulation proceeded with boundary condition and other data being exchanged between the machines via the ATM net. The system was successfully demonstrated to UCAR management and represented the first instance of distributed supercomputing between NCAR's Boulder laboratories. This system opens up the possibility of real-time distributed mesoscale weather prediction using a collection of computing facilities across the U.S. among participating institutions. By distributing the computational workload between various sites, the system allows a major increase in computing power for truly high-resolution (1-5 km) mesoscale weather prediction.
Over the last year, Andrew Crook (joint appointment with RAP) and Juanzhen Sun examined the physical mechanisms that underlie single Doppler velocity retrieval (SDVR) methods. Attention was focused on two SDVR techniques, the Gal Chen-Liou method and the adjoint method. Simulated data of a collapsing cold pool were used to compare the two techniques. Crook and Sun interacted with researchers from the University of Oklahoma, Center for Analysis and Prediction of Storms (CAPS), and Pennsylvania State University.
The Gal Chen-Liou technique is premised on the assumption that the buoyancy in a flow can be retrieved from knowledge of one velocity component. The retrieved buoyancy is then used in a numerical model to drive motion in the cross-beam direction. After 20 iterations of the Gal Chen-Liou technique, approximately 70% of the buoyancy signal was retrieved, while only 30% of the cross-beam flow was recovered.
The adjoint method retrieves cross-beam velocity not only by the accelerative mechanism described above, but also by tracking the radial velocity field with time. Results have shown that to effectively utilize accelerative retrieval it is necessary to apply the adjoint method over successive time windows using the retrieval from the previous cycle as a first guess and background field in the next retrieval. After three successive applications of the adjoint technique, the performance is now significantly improved, with 83% of the buoyancy field retrieved and 70% of the cross-beam velocity recovered. Further improvement can be gained if a tracer, such as clear-air reflectivity, exists in the flow and can be tracked.