by Bob Henson
NCAR’s Yongsheng Chen and Richard Rotunno worked with a team of colleagues to study an idealized hurricane at four resolutions (see labels below.) At the tightest resolution of 62 meters, turbulence broke the smooth eyewall into a set of small convective cells. (Photo by Carlye Calvin, illustration courtesy Rotunno and Chen.)
If you zoom too far into a digital photograph, what was once a recognizable scene can morph into a seemingly chaotic collection of pixels. NCAR modelers Richard Rotunno and Yongsheng Chen had a similar experience with one of the most detailed simulations of a hurricane ever produced.
“We were trying to get at details that aren’t usually expected in numerical prediction models,” says Rotunno. Indeed, it took a special computing allocation from NCAR’s Breakthrough Science program, and a customized version of the multiagency Weather Research and Forecasting model, to make their project happen. Also involved were NCAR’s Wei Wang, Christopher Davis, Jimy Dudhia, and Gregory Holland. “It was really a group effort,” says Rotunno. “We met at 5:00 p.m. every day for several months to look at the previous day’s simulations and figure out what to do next.”
Starting at 15 kilometers (9.2 miles) between data points, the group tried out six progressively tighter grids, ending up with a mere 62 meters (200 feet)—only about half a city block—between points. They watched intently for the appearance of a key player: turbulence. The turbulent eddies that swirl around towering storm clouds are too small for most models to depict, but they play a huge role in modulating the intensity of the showers and thunderstorms (convection) that feed into hurricanes.
With each new level of detail, the modelers saw their idealized hurricane strengthen. Would turbulence finally appear and put on the brakes? It wasn’t until the scientists moved to the tightest grid of their experiment, the 62-m resolution, that the smooth eyewall of the hurricane suddenly disintegrated into dozens of tiny, shard-like convective cells (see graphic).
Far from being a failure, the collapse of this pretty picture was a triumph. The resulting hurricane looked much more like the real thing, convective warts and all. And though the wind gusts at this resolution were the highest of all their simulations, the average winds actually dropped off—a clear sign of turbulence at work.
The computing power won’t be in place for years to include this kind of detail in everyday hurricane models. But the tests being conducted by researchers at NCAR and elsewhere are helping find the sweet spots of resolution, where important phenomena such as thunderstorms and turbulence can be included at a minimum of computational cost. (Each doubling of a model’s horizontal resolution requires an eightfold increase in computer time.) Their work is also helping gauge the proper balance between boosting model resolution and tackling other problems that stand in the way of better hurricane forecasts.
“We suspect that better resolution helps, but we don’t know its benefits relative to other things,” says Rotunno. “Do we want to put resources into higher resolution or into some other feature, like a more sophisticated initialization system or more ensemble members? There are some real tradeoffs.”
Where, and how intense?
Hurricane prediction has had its strengths and weaknesses for some time. On the plus side, models can now nail down the tracks of most hurricanes with reasonable success, since the steering flow is normally governed by large-scale weather features that are well forecasted. In recent years, the average 48-hour track error from NOAA’s National Hurricane Center has been just under 160 km (100 mi). For the infamous Hurricane Katrina, the NHC’s 48-hour forecast was only 79 km (49 mi) miles off.
Intensity forecasts are another matter. A hurricane’s core can be tiny—often smaller than a Midwest thunderstorm complex—and many subtle features affect its growth. Even after years of research, today’s forecasts have little skill at predicting whether a tropical cyclone in seemingly prime conditions will strengthen or weaken. This year, Hurricane Dolly rapidly intensified to Category 2 status on the Saffir-Simpson scale on 23 July, just before striking Texas. Only a few hours earlier, none of the major models had predicted Dolly to strengthen beyond Category 1.
Naomi Surgi (above, from the NOAA Environmental Modeling Center) is coordinating development and refinement of NOAA’s Hurricane WRF model. HWRF is part of the guidance package used by forecasters such as Lixion Avila (left) at the National Hurricane Center. (above photo courtesy NOAA; left photo by Bob Henson.)
Some help has arrived in the forecaster’s toolbox with NOAA’s adaptation of the Weather Research and Forecasting model. Dubbed the Hurricane WRF (HWRF), the model entered operational use in 2007 after years of testing and development at NOAA’s Environmental Modeling Center (EMC). HWRF proved its mettle as the most accurate of the operational models for 2007’s Hurricane Dean at two critical points: when Dean struck Mexico’s Yucatán Peninsula as a Category 5 storm and the state of Veracruz as a Category 2.
Upgrades this year to HWRF’s initialization and physics schemes promise to cut its average intensity error by as much as a third, says NCEP’s Naomi Surgi. “There are many difficult problems ahead, but we’re making progress,” says Surgi. “More progress will come through close collaboration between research and operational modelers invested in producing the best possible guidance.” Starting at the end of the 2008 hurricane season, Surgi will spend a year at the NCAR-based Developmental Testbed Center, where variations of WRF and other models are explored in a setting that serves both research and operational needs.
Resolving the situation
Megan Gentry (North Carolina State University) presented her analysis of hurricane-model resolution at the American Meteorological Society’s conference on tropical meteorology, held last April in Orlando. (Photo by Bob Henson.)
What happens when resolution goes from HWRF’s 9 km down to 1 or 2 km? “It’s called a ‘no man’s land’ in the literature,” says Rotunno. The convection that spirals into hurricanes serves as the cyclone’s engine, yet many convective cells are only 2 to 6 km wide. Thus, they can’t be simulated directly in most of today’s operational models, including HWRF.
Since 2005, researchers have had a powerful tool at hand for exploring the role of convection in hurricanes. The Advanced Hurricane-research WRF (or AHW)—specially tailored for studying tropical cyclones—is an offshoot of NCAR’s Advanced Research WRF (ARW), which is designed to study a range of mesoscale weather features using a different set of model physics than the operational versions of WRF.
In real-time forecast studies conducted over the past several years, AHW accommodated resolutions as tight as 1.33-km (0.83 mi), which is enough to explicitly portray showers and thunderstorms. Facing limited resources, NCAR put the AHW hurricane experiments on hold this year in order to contribute to NOAA’s 10-year Hurricane Forecast Improvement Project (see “On the Web”).
The idea behind AHW isn’t to replicate each bit of convection that occurs in an actual hurricane, but rather to simulate more realistic convection—including the kind of sudden blow-ups that can help a hurricane rapidly strengthen. “If your model can’t pick up on that eruption of convection, you may not be able to simulate hurricane intensity,” says NCAR’s George Bryan.
In general, higher resolution leads to stronger hurricanes in models, as the better-resolved convection is able to organize and intensify more readily. At North Carolina State University, graduate student Megan Gentry and adviser Gary Lackmann used ARW to explore the effect of six different resolutions, ranging from 8 to 1 km, on simulating Hurricane Ivan from September 2004. Gentry found that tightening the grid spacing to 2 km produced a hurricane that was as much as a full category stronger than at 4 km.
In an upcoming paper, Bryan and Rotunno find that models in the “no man’s land” range, such as AHW—the ones that explicitly include thunderstorms but not turbulence—can overestimate hurricane strength by as much as 50%. “Turbulence is a strong determinant of hurricane intensity,” says Bryan, “but the forecast models can’t handle it very well. It’s an important process, and we know almost nothing about how to deal with it.”
Hundreds of times more computing power may be needed before routine hurricane forecasts can get the benefit of incorporating turbulence. In a study that preceded the NCAR work led by Rotunno and Chen, Bryan teamed with John Wyngaard and Michael Fritsch (Pennsylvania State University) to explore how thunderstorms evolved for model resolutions that ranged from 1 km to 125 meters. The team’s 2003 paper in Monthly Weather Review reported that the tightest resolution produced the richest, most satisfying depictions of thunderstorms, with turbulence helping sculpt the convection in a realistic way.
The shape of hurricane models
As the NCAR modeling team found, a cyber-hurricane need not look elegant to be realistic. Models that can explicitly depict thunderstorms often produce the asymmetric, wobbly-looking eyes found in many real hurricanes. Moreover, models such as AHW have shown the ability to capture eyewall replacement cycles, in which one eye tightens and collapses while a larger one develops and eventually dominates. This process, which takes a day or two to unfold, often causes a dip in intensity followed by restrengthening, so it’s crucial to a skillful forecast.
In her simulations of Hurricane Katrina at 1.33-km resolution using NCAR’s Advanced Hurricane-research WRF model, Kristen Corbosiero (above, from UCLA) found a triangular eyewall, as shown here in a map of precipitable water. Each of the triangle’s corners, where wind speeds were strongest, took about 40 minutes to rotate around the eye of the model-simulated Katrina. Although Katrina’s actual eyewall was not triangular, the results are proving useful in analyzing how high-resolution models deal with hurricanes. (Photo by Bob Henson; illustration courtesy Kristen Corbosiero.)
In collaboration with NCAR modelers, Kristen Corbosiero (University of California, Los Angeles) found that the 1.33 km AHW produced an abundance of unusually shaped eyewalls, including many triangles (see graphic), whereas actual hurricane eyewalls are more likely to resemble squares or ellipses. “Triangular features are rare in nature. This is one of the mysteries of these high-resolution simulations,” says Corbosiero. She and UCLA colleague Robert Fovell are exploring how model parameterizations may be involved and how the asymmetric features relate to storm intensity.
Over the last decade, forecasters have also learned the importance of high ocean heat content—warm water that extends unusually deep—in supporting the rapid growth of hurricanes. Now researchers are striving to couple the newest hurricane models with high-quality ocean models, and they’re looking for better ways to incorporate data from satellite-borne altimeters that measure ocean height (an index of water that’s warming and thus expanding upward).
In the June issue of Monthly Weather Review, NCAR’s Christopher Davis and colleagues show that an AHW simulation of Katrina was noticeably improved with the help of a mixed-layer ocean model. This relatively simple model provides most of the oceanic cooling feedback of importance to hurricane intensity forecasting without the computing overheads of a full-fledged ocean model. “This is one of the advantages of conducting research simulations in real time,” says NCAR’s Holland. “We can try out innovative ways of arriving at the best possible result within the available computer resources.”
NOAA’s HWRF, which is coupled to the Princeton Ocean Model, already includes upper-ocean data gathered by probes dropped from hurricane-hunter aircraft. Next year the model will be coupled to an advanced ocean model developed at the University of Miami, with the ability to assimilate a continuous feed of ocean data and provide 3-D analysis of key features, including warm-core eddies—implicated in the rapid growth of 2005’s Katrina and Rita—and their cold-core counterparts. Further improvements by 2010 will help the model depict battering waves, wave-driven sea spray, and other aspects of air-sea interaction, and by 2012 HWRF will be coupled to an advanced storm-surge model. “We’re building this system component by component to address the complete problem of coastal inundation,” says Surgi.
For now, hurricane forecasters are keeping their eyes on the big picture, using ever-growing ensembles of model guidance to forge a consensus rather than taking any particular model’s forecast too literally. On EMC’s drawing board is a plan to blend five to six high-resolution hurricane models into the current mix of three operational models, leading to a unified package to guide forecasters.
“We recognize the use of ensembles as a powerful method for improving hurricane forecasting,” said National Hurricane Center director Bill Read. “We’re hopeful the plans and efforts of NOAA and NCAR scientists to develop a high-resolution multi-model ensemble will translate to better intensity computer guidance for our forecasters and better forecasts for the public.”
As for the high-resolution frontier, Davis points out that, while many realistic features are being simulated, “these structures have yet to be quantified objectively.” Such a task may demand new measures of forecast skill, he says.
Much like Davis, Chen is intrigued but cautious: “We still have many issues to resolve before we can trust these high-resolution simulations.”