ESRL Quarterly Newsletter - Fall 2009

Storm Spoiler

ESRL’s FIM model excels during summertime hurricane trials

Hurricane Bill couldn’t outrun FIM. The ESRL developed forecast model kept pace with that storm and others this summer in a series of supercomputer-based hurricane forecast experiments.

“The results are preliminary, but the track forecasts seem to be looking better when the model is run at higher resolution (30 km vs. 15 km vs. 10 km),” said ESRL’s Stan Benjamin, chief of the Assimilation and Modeling Branch of ESRL’s Global Systems Division. “At higher resolution, we are seeing intensity improvements, too, based on this season’s tropical storms so far,” Benjamin added.

Forecast tracks by the FIM global model.

Forecast tracks by the FIM global model and global ensemble, initialized at 0000 GMT Aug. 20. The black track is Hurricane Bill’s observed track. White tracks are from the 20-member, 30-km ensemble; magenta and green are 30- km deterministic runs with different initial conditions (EnKF vs. GFS); yellow and red tracks are from the 15-km deterministic model with the different initial conditions; and blue is from the 10-km deterministic run using the EnKF initial condition.

He and colleagues from across NOAA, supported by the Hurricane Forecast Improvement Program, HFIP, used TACC supercomputers to run FIM at high resolutions this summer. Although forecasters have improved hurricane track errors by about 50 percent during the last 20 years, HFIP has set a goal of further, significant improvements in track and intensity forecasts and forecast lead time.

It’s hard to spin up a realistic, intense hurricane in lower-resolution models, such as the operational Global Forecast System (GFS), which usually runs at 40 km. The storms alter the wind and other systems in which they’re embedded, and those changes can then push around the storm itself. Researchers calculate they need resolutions of 5-15 km to perform this kind of modeling, and those resolutions require intense computing power (shifting FIM from 30-km resolution to 10-km resolution required 27 times more computing power).

So this summer, Benjamin and his colleagues used more than 12 million processor hours at TACC to run hurricane forecasts on FIM at 30-, 15-, and 10-km resolution, to study the effect of resolution on forecast accuracy. The team is also comparing the effectiveness of two methods that could be used to initialize the models. The first is a traditional data assimilation system, used operationally with GFS, that ingests real-time meteorological data from a broad region around developing hurricanes. The second is an experimental system, developed at ESRL, which treats those data first with a statistical technique called the ensemble Kalman filter, EnKF. Initial data from those trials already suggest that both FIM and the operational GFS can deliver better forecasts using the EnKF.

See it on the Web