Slashdot Mirror
Supercomputers Help Researchers Improve Severe Hail Storm Forecasts (nsf.gov)
aarondubrow writes: Researchers working on the Severe Hail Analysis, Representation and Prediction (SHARP) project at the University of Oklahoma used the Stampede supercomputer to gain a better understanding of the conditions that cause severe hail to form, and to produce hail forecasts with far greater accuracy than those currently used operationally. The model the team used is six times more resolved that the National Weather Service's highest-resolution forecasts and applies machine learning algorithms to improve its predictions. The researchers will publish their results in an upcoming issue of the American Meteorological Society journal Weather and Forecasting.
And the thing is, Ike could have been so much worse for Galveston. If landfall had been ~10 kilometers to the southwest (which in hurricane terms is just a wobble), he would have done to Galveston what he did to the Bolivar Peninsula.
I was so mad with the mayor of Galveston, constantly playing down the building storm until the last minute out of fear of driving away tourist money. The NHC was taking the storm very seriously and giving warnings about its size, its growth potential and the potential range of its impact location, and then the mayor would come on and say, no, no, it's going to hit way south of us, and even if it did hit it'll only be a little category 1 or 2...
As bad as the damage was, they could have lost half the bloody island (along with all of the people who waited too long to leave due to listening to idiots like their mayor) had it made landfall just a bit further to the southwest. The northeast quadrant is the dangerous part of an Atlantic-basin hurricane. A more southwesterly track would have added nearly a meter to the storm surge and significantly increased winds and wave heights. And it would have also attacked the island on less well defended areas.
Hourglass says she knows a kid in Iowa who grows up to be president.
The article is fairly light on the details of what they're doing. It's also deceiving.
Warn-on-forecasting is highly unlikely to result in severe thunderstorm warnings issued hours in advance. Right now, warnings are generally issued when any two of these three things occur: 1) the atmosphere is favorable for severe or tornadic storms; 2) radar indicates that severe weather is probably occurring; 3) storm spotter reports indicate that severe weather is occurring. If spotter reports are credible enough, warnings may be issued on the basis of those even if the other two conditions are marginal. Warnings are issued based on observations of something that's happening now. Warn-on-forecasting will involve issuing warnings because numerical models of a storm indicate there is a high probability a thunderstorm will become severe or tornadic in the future. These warnings will still be based on actual storms that have already formed and are likely to be for lead times of an hour or two.
We already have predictions for when storms are likely to be severe hours or days in advance. The Storm Prediction Center issues convective outlooks as much as eight days in advance. They also issue severe thunderstorm and tornado watches, which are hours in advance. They indicate areas where there's a high probability of severe weather, but they're not forecasts for individual storms. They're more general in nature.
What's really going on is that researchers have data mined the characteristics of a very large number of thunderstorms to determine what characteristics are the best indicators of hail size. They've developed another algorithm for identifying and tracking individual thunderstorms, so they can track those characteristics over time. It seems like the algorithm may be an improvement over prior storm identification and tracking algorithms, though a huge number of these algorithms exist. They're using the algorithm to track storms in both observations and forecasts that come out of numerical models. They've also improved the grid spacing of the models, going from 3 km horizontal (the resolution of the HRRR model) to 500 m horizontal. Presumably they've added more vertical levels. When you increase the resolution, you also need shorter time steps in the model, otherwise you'll get what are called CFL errors and the model won't run (or won't run properly). I believe dt for the HRRR is around 18 seconds, so you'd probably have to lower dt to 3 seconds (maybe lower) for a 500 m model. Also, instead of running the model once, they're running an ensemble, meaning that they have 50 or 100 different model simulations with slightly different initial conditions so they can determine the probabilities.
What's missing from this are three things, though those researchers may be working on them anyway. One is that forecasts are only as good as the observations going into the model and its initial conditions. Getting good initial conditions is a challenge for thunderstorms because you need good observations of the storm structure (radar can help with this, but more observations are needed) and good data of the storm environment. We have lots of surface observations, but few observations above the surface. The first 1,500 meters or so above ground level are very important in how strong or tornadic storms are. We don't have a lot of observations above the surface in this area, but they're important. We also need better data assimilation techniques so those observations translate to better initial conditions for our models. Finally, we can improve models in ways other than data assimilation and improved resolution. We don't directly simulate hail in models; there simply isn't the computing power to simulate individual hydrometeors (e.g., hail stones, raindrops, ice crystals, and cloud droplets). We parameterize those using microphysics schemes in models. We've improved the microphysics schemes a lot over time, adding more classes of hydrometers and adding the ability to predict their size distribution and other data about them. Ge