Starting in June 2013, scientists around the world began downloading software that can simulate and forecast weather on a global basis. The debut of the Model for Prediction across Scales (MPAS) marked the first time that researchers everywhere could freely gain access to global-scale weather modeling tools that offer a level of detail previously available only in models spanning a particular region. The model is being jointly developed by NCAR and the U.S. Department of Energy’s Los Alamos National Laboratory, with NCAR focusing on the atmospheric component and LANL the oceanic component.
MPAS follows in the highly successful footsteps of the Weather Research and Forecasting model (WRF), which was created by a multiagency partnership in the late 1990s and early 2000s. NCAR’s version of this model, the Advanced Research WRF (ARW), has been used as an open-access tool by more than 2,000 scientists in over 150 nations. NCAR provides free, downloadable ARW software as well workshops and tutorials that focus on the model and how to get the most out of it. Many hundreds of research papers have drawn on ARW simulations. The model has also proven its mettle as a forecasting tool during field campaigns focused on hurricanes, tornadic thunderstorms, and other types of extreme weather.
Despite its high value and widespread use, WRF faces an obstacle common to many other weather models: the tyranny of the latitude-longitude grid. Such models slice the atmosphere into blocks, or grid cells, that are typically many kilometers wide and a few hundred meters tall, bounded by north-south and east-west lines. However, because longitudinal lines converge toward the North and South Poles, the grid cells are not quite rectangular. At higher latitudes, they become increasingly narrow, and special computational techniques are needed to keep the model’s atmosphere working in a realistic way.
MPAS gets around this roadblock through the use of what is known as an unstructured Voronoi mesh (see illustration). Instead of a grid with rectangles that taper toward the poles, the MPAS mesh features a latticework of shapes—mostly hexagons, but with a few five- and seven-sided cells—that intersect in a seamless way across the entire globe. The resulting structure bears some resemblance to the shape of C60, the buckminsterfullerene (or “buckyball”) molecule. But while that molecule has just 32 faces (as does a standard soccer ball), the MPAS grid can have millions of cells.
Along with eliminating the problems caused by polar convergence, the meshed grid of MPAS has another benefit: it can be easily telescoped to provide higher resolution in those regions where extra detail is desired. Traditional model grids can also be nested in this way, but the abrupt transitions along the sharp edges of a nest can be computationally problematic. In MPAS’s variable mesh, the transition from coarse to fine resolution is seamless.
One of the most in-depth projects for MPAS thus far is a set of runs carried out after the installation of NCAR’s Yellowstone supercomputer in late 2012 at the new NCAR-Wyoming Supercomputing Center. These runs were part of NCAR’s Accelerated Scientific Discovery (ASD) program, which allows a select set of computing-intense research projects to be carried out before a newly installed supercomputer is made available for more general use.
Under the ASD project, MPAS was run in retrospective mode. The model produced 10-day forecasts for two periods from 2010, allowing the forecasts to be compared to actual weather observations and satellite images. In these tests, the model’s variable mesh narrowed from 60 kilometers to as fine as 3 km in areas of special interest, such as where severe weather might be expected. Grids cells larger than about 10 km are too coarse to simulate individual showers and thunderstorms, so the presence of storms must be specified, or parameterized—somewhat like predicting the course of a football game based on general statistics from past games. Even a 3-km grid is still too coarse to track individual cloud elements, but it does allow for thunderstorms to form, grow, and die over time—essentially simulating each play of the football game, rather than relying on statistical clues.
The abilities of MPAS were especially evident in one of the retrospective ASD forecasts, issued from starting-point conditions for the evening of October 22, 2010. The model predicted that four days later, on the afternoon and evening of the 26th, a broken line of supercell thunderstorms would move through the Ohio Valley. That forecast closely matched reality, with supercells marching across the region and producing dozens of tornadoes from Alabama to Ohio.
A major test of the model was carried out during the 2013 Atlantic hurricane season. Ten-day forecasts were generated once a day from 1 August to 20 September using two versions of MPAS. One version featured a variable mesh that transitioned from 60-km resolution across most of the world to 15 km across the tropical Atlantic and neighboring land areas. The other version spanned the entire globe at 15-km resolution. Although it provided more overall detail, the full 15-km version also required five times more computing time and expense. One of the key aspects of this project was to see if the less expensive variable mesh would perform as well as the uniform mesh for tropical cyclones.
The answer came from the Eastern as well as the Western Hemisphere. In 2013, the North Atlantic experienced an unusually quiet season for tropical cyclones—there were only two hurricanes, the least since 1982—so the researchers had only a handful of cases to evaluate against the MPAS forecasts. However, researchers at the Taiwan Typhoon and Flood Research Institute ran a companion study from 1 August to 31 November, centering the variable-resolution MPAS grid over Taiwan. This provided a larger set of test data, since the northwest Pacific experienced a much busier tropical season in 2013, even apart from the catastrophic Supertyphoon Haiyan.
Together, the Atlantic and Pacific studies showed that accuracy didn’t suffer in a significant way when the less computationally costly grid was used. In some cases, forecasts up to six days in advance were able to capture specific tropical cyclones reasonably well in both versions of MPAS (see graphic).
The versatility of MPAS makes it a potential candidate for what’s known as a unified model, one that could be used economically in place of several other models at an operational forecasting center. Even finer-scale variable meshes will be explored in the near future. A major question is how MPAS will perform at resolutions below 1 km; at this level of detail, individual cloud features can be resolved, such as the air parcels entrained into a building thunderstorm. Another goal is to develop new techniques for bringing information on current weather into MPAS, in tandem with NCAR’s Data Assimilation Research Testbed.