Forecasts from Numerical Weather Prediction (NWP) models provide one of the primary tools meteorologists use to produce weather forecasts. Historically, running NWP models has required vast computing resources to complete weather forecasts in a timely fashion. Until recently, running such NWP simulations quickly at a high enough resolution to capture mesoscale features (such as tight temperature gradients, mountain/valley flows, and mesoscale precipitation banding features within midlatitude cyclones) required the purchase of a supercomputing cluster. However, the rise of cloud computing technologies has removed that barrier. Now, companies like Google, Amazon, and Microsoft provide the required computing resources at a per use cost to the public and academic communities.
This fall, Dr. Siuta’s junior-level Analysis and Forecasting 1 class won a Google Cloud for Education grant, which allowed them to use Google’s Cloud Platform to run their own NWP model simulations using the state-of-the-art Weather Research and Forecasting (WRF) model. Dr. Siuta had previously co-authored an article on the Viability of Cloud Computing for Real-Time Numerical Weather Prediction in the journal Weather and Forecasting.
Students learned the basic components of a NWP model, weather model limitations, and how adjusting model physics can lead to different forecast outcomes through simulating a high-impact weather event of their choice. Students ran cases covering Superstorm Sandy (October 29-30, 2012), the February 8-10, 2013 nor’easter, the March 7-9, 2018 nor’easter, the March 1-3, 2018 nor’easter, the January 4-6, 2018 nor’easter, and the October 2017 New England wind storm. Unidata’s Integrated Data Viewer was used to visualize the WRF output.
The grant from Google was sufficient to cover the simultaneous use of 336 virtual computing cores, so that students could run each of these cases with model resolutions matching the standards of today’s national model centers — down to the 3-5 km scale over the entire northeastern US.
Key findings by the students are summarized below.
Acknowledgements: We thank the Google Cloud Platform for providing the funds through the GCP Education Grants program for our class.
Case Study 1: October 29-30, 2017 wind storm
WRF model sea-level pressure (left panel) and wind speed (right panel) output for the 29-30 October 2017 wind storm. This case study was chosen by Alex DaSilva and Nick Ferrando-Boucher, who varied the WRF planetary boundary layer scheme to see the effect on the strength of the low-pressure center and magnitude of the wind speeds in Vermont. They found that a non-local-mixing boundary-layer scheme provided a better forecast than a local-mixing only boundary-layer scheme for the wind speeds observed at Northern Vermont University-Lyndon during the event.
Case Study 2: March 7-9, 2018 Nor’easter
A comparison of WRF model forecast snow depth (using 10:1 ratio) to that of observations in southern New England for the 7-9 March 2018 Nor’easter (Quinn). This case study was run by Sarah-Ellen Calise and Lauren Cornell, who varied WRF cloud microphysics schemes to see the impact each scheme had on snowfall forecasts. D01 are forecasts for a 15-km outer nest while D02 are forecasts for a 5-km inner nest. Sarah-Ellen and Lauren found that in areas closer to the coast (e.g., Providence, RI), the microphysics scheme had a substantial impact on the location of the rain/snow line and overall snow amounts in the area.
Case Study 3: Superstorm Sandy (October 27-30, 2012)
Jonathan Hutchinson, Taylor Leitch, and Lillie Farrell varied the planetary boundary layer schemes in the WRF model to see the impact on the development of Superstorm Sandy. Shown here is one of their WRF simulations predicting landfall on October 28 along the coast of New Jersey as Sandy is undergoing a transition from a tropical to extra-tropical system. The group found ~10-mb sea-level pressure difference between their runs at Atlantic City, NJ caused by differences in forecast track. Landfall varied between Sandy Hook and Cape May, NJ depending on the planetary boundary layer scheme that was used.
Case Study 4: February 8-10, 2013 Nor’easter
Students John Drugan, Radek Przygodzki, and Alex Doone ran the February 2013 nor’easter, which resulted in heavy, wet snow and 700,000 people losing power in parts of New England. John, Radek, and Alex varied cloud microphysics scheme and determined that the microphysics choice had a distinct impact on the location of a mesoscale precipitation band forecast to occur near the eastern tip of Long Island. Depending on the scheme, this band shifted to the east or west, leading to the highest forecast storm totals shifter slightly towards (right panel) or away (left panel) from population centers.
Case Study 5: January 4-6, 2018 Nor’easter
Kelsey Emery and Dan Carneiro simulated the January 2018 nor’easter because they were personally affected by the storm, which prevented them making it to the national American Meteorological Society meeting due to disrupted flights. The storm hammered southern New England with 1-2 feet of snow and winds gusting over 50 mph. Shown below are Dan and Kelsey’s WRF simulations of the low tracking off the Massachusetts coast.
Case Study 6: March 1-3, 2018 Nor’easter
A comparison of model forecast soundings produced by varying the boundary layer choice during the March 1-3, 2018 nor’easter. This nor’easter left close to 2 million people without power in the northeastern US due to wet snow and high winds. The far right graphic shows the corresponding observed sounding at Albany, NY taken 0000 UTC 3 March 2018. Rosemary Webb and Sarah Sickles found that varying the boundary layer choice affected the ability of the model to depict the vertical profile of the atmosphere at Albany, NY.