High-performance computing helps researchers predict air pollution events

Urban air pollution is a major concern for Santiago, Chile, home to nearly 6 million people. The city lies in a basin surrounded by the Andes Mountains, and vehicle and industrial emissions tend to linger – especially in May June, and July when weather conditions make the air more stagnant.

In an effort to protect public health, the local government uses predictions to declare air pollution episodes. If an episode is anticipated, officials impose temporary restrictions on activities such as driving and manufacturing to reduce the high levels of smog. If predictions indicate that the pollution could reach dangerous levels, precautionary measures are even more stringent – like keeping children home from school and advising people with breathing problems to stay indoors.

“Forecasts for these episodes are conducted 24 hours in advance, but by waiting until an episode is imminent, the efforts to prevent the episode don’t do much good,” says University of Iowa graduate student Pablo Saide, a native of Santiago. “The hospitals are so crowded full of people with asthma and kids. And people develop cancer and long-term health problems because of the pollution.”

Motivated by the predicament in his home country, Saide came to the United States to pursue a doctorate in environmental engineering. With a team of researchers at the UI, he is working to address the “too little too late” situation by developing better pollution prediction models. Their work is possible because of high-performance computing (HPC) resources that the university offers through Information Technology Services (ITS), along with grants from NASA and the National Science Foundation.

Already, engineers at the UI Center for Global and Regional Environmental Research (CGRER) have developed computer model simulations that can forecast air pollution up to three days in advance. Now, they are comparing their predictions to those of a local observational network in Santiago to see if the new model is more accurate.

Data for the simulations comes from measuring stations around the city that monitor how the plume moves, horizontally and vertically. The air-quality measurements, along with meteorological and forecasting data, feed into models that can be quickly processed by the UI’s Helium computing cluster. This enables scientists to quickly generate high-resolution simulations to predict pollution.

“We could create simulations like these using a laptop, but the resolution wouldn’t be nearly as high, and it would take ten times longer,” Saide says. “You cannot take a week to develop a forecast for the next day. You need it in hours. High-performance computing allows us to do that.”

In a separate but related project, Saide and colleagues are investigating whether their model could also produce more accurate predictions for climate change.

An important process that affects climate change occurs when pollution particles interact with clouds, changing the clouds’ properties by making them more efficient in reflecting solar radiation. Most of the margin of error in existing climate change models is a result of uncertainty about how the particles affect the clouds.

To gain a better understanding of that relationship, scientists are studying a location off the coast of Chile that is cloudy 80 to 90 percent of the time. Emissions from the copper smelting industry and other anthropogenic activities affect the clouds. A measurement campaign in 2008 gathered data on the clouds and particles from instruments inland and on ships, aircraft, and satellites.  The researchers compare that data against their simulations in order to validate the model.

“Our job, after we get the observational data, is to run our model to see how well it represents the actual situation,” Saide says. “We test our model against satellite data to see how accurate it is. When we see differences, we tweak the model configuration and run it again to keep refining and improving it. Once the model is able to represent the particle-cloud interactions, meaningful studies on the effect of anthropogenic pollution over clouds and climate can be made.”

Because of the massive volumes of data involved, the research team uses 250 processors in parallel to run a single simulation.

“A desktop computer usually has eight processors, so using Helium is the equivalent of running about 30 desktop machines,” Saide says. “It still takes two weeks to run the model with HPC, but I can’t even imagine trying to do it with a single machine. It would probably take a year.”


The image shows Pablo Saide, as well as an image of two MODIS-Aqua products for Oct. 17, 2008, over the persistent Southeast Pacific stratocumulus deck off the coasts of Chile and Peru. The background is a true color image of the clouds at 250 m resolution, with an iridescent overlay showing a 1km resolution cloud droplet number (Nd) retrieval for the same overpass. Red, green, and blue colors show high (~1000 #/cm3), medium (~100 #/cm3) and low (~10 #/cm3) Nd values. The east to west Nd gradient is produced mainly by anthropogenic aerosols, which influence cloud properties.