How long can you hold your breath?
By Christina Daniilidi, computer science
Everyone enjoys going out for fresh air. “But the air outside may not be as fresh as we hope — due to air pollution,” warns Virginia Tech computer scientist Adrian Sandu. “And, since breathing is inevitable, there is no way to avoid inhaling bad air.”
Wouldn’t you like to know what you are inhaling, and have accurate detailed forecasts of air quality? Sandu, an associate professor and head of the Computational Science Laboratory, is creating models to answer these questions, whether you are in Chicago or an industrialized area of China. His sophisticated models of air quality are also helping policy makers clean up our air.
Almost a third of the population lives in areas where air pollution levels exceed U.S. Environmental Protection Agency (EPA) health-based standards for air quality, says Sandu.
“Tens of thousands of people die each year as a direct result of exposure to high levels of air pollution; many more suffer adverse health impacts.”
Crop yields and forest productivity are also adversely impacted by air pollution. In 2004, the U.S. Department of Energy estimated that “the total cost to the U.S. economy of issues related to the environment exceeds $100 billion per year, and the decisions being made are far from optimal.”
“Although the idea of combining model predictions and observations is common sense, its full realization is extremely challenging, and is the focus of intense research at the cutting edge of scientific computing.”
The major pollutant usually associated with poor air quality is ozone. While ozone in the stratosphere shields life from UV radiation, elevated ozone at ground level damages living tissue of plants and animals.
Other important air pollutants include nitrogen oxides, volatile organic compounds, carbon monoxide, and sulfur dioxide. Fine particles from burning vegetation, sea-salt spray, blowing dust, and volcanic ash are also important pollutants.
Sound policy and management strategies to protect society against serious environmental threats require a deep understanding of the processes that control pollution formation and distribution. This scientific understanding is built through experiment, analysis, and modeling.
Experimental data are collected in laboratory investigations, during intensive field studies, and from extensive observation networks. Analytical studies propose mathematical equations to describe the physical and chemical changes of the atmosphere. Comprehensive computer models simulate the evolution of the chemically altered atmosphere.
The development of accurate computer models remains a major challenge due to the complex processes occurring at widely different scales and their strong coupling across scales.
Spatial scales range from the molecular, such as the formation of new particles, to the continental, which are associated with long-range transport of pollutants. Temporal scales range from micro-seconds, which is all some chemical reactions require, to years, as in the case of such long-lived trace gases as methane and carbon dioxide.
Building “in-silico” atmosphere
The Computational Science Laboratory (CSL) is developing numerical methods and software to create modern computer models to simulate the polluted atmosphere. Such models account for the complex interactions between many processes, including the emissions of pollutants from natural and man-made sources, chemical transformations, and long-range transport. Global chemical transport models track the time evolution of tens to hundreds of millions of variables and pose interesting and difficult problems from computational and mathematical points of view.
“Numerical methods used to solve the chemical and physical evolution equations need to be highly efficient, accurate, and stable,” says Sandu. “They should preserve the important physical characteristics of the atmosphere. For example, mass should not be artificially created or destroyed by computational processes.”
Sandu and CSL develop highly efficient methods for solving the atmospheric chemical rate equations. One of the group’s methods, a “solver” called Ros3, has been adopted by the EPA’s congestion mitigation and air quality (CMAQ) model, the agency’s main tool for national air quality forecasts, research, and policy development. A solver is an algorithm that solves equations. The Ros3 solver helps the model solve equations required to account for trace gases in the atmosphere.
In addition to elegant math, the CSL group adopts novel approaches to overcome such traditional modeling challenges as resolution, or scale, and computation speed.
For example, CSL has developed high-accuracy numerical methods for simulating aerosols — tiny particles suspended in the atmosphere, which are often harmful when inhaled. Aerosols undergo physical, chemical, and thermodynamic changes that are mathematically described by a set of challenging integral-partial-differential equations.
Insufficient geographic grid resolution is known to cause considerable loss of accuracy in air quality models. Sandu’s group is investigating the use of adaptive grids to automatically place additional resolution only where it is needed, thus improving accuracy without increasing the computational effort. For example, in a simulation of East Asia, this “smart” code has automatically chosen a 10-kilometer (6.2-mile) resolution above the industrial areas in China, but a 160-kilometer (100-mile) resolution above the Pacific Ocean.
The CSL team is also developing sophisticated solvers that use different time increments in different parts of the atmosphere. This “multi-rate” approach has provable mathematical properties, like high accuracy, conservation, and stability. “It avoids the necessity of accounting for small time increments across the globe, for example, to accurately resolve the pollution in the Chicago area,” Sandu says.
A new concept that his group promotes is to develop systems that can automatically write complex model software in an efficient, error-free manner. One example is the Kinetic PreProcessor, which writes chemical simulation code. It is being used by academic, research, and industry groups in several countries, including the United States, Germany, the Netherlands, France, Italy, Ireland, and Denmark.
Harnessing the power of large-scale parallel computers is essential to manage the computational complexity of air quality simulations, and Sandu’s group is developing tools to facilitate the parallel implementation of these models. An example would be the communication library, PAQMSG (parallel air quality models on structured grids), which splits the atmosphere into small subdomains and feeds the data chunks to many different processors, which then simultaneously perform local simulations. The processors periodically exchange information, producing results from the distributed computations that are identical to those that would be obtained if the full atmosphere was simulated on a single computer – except faster.
The group is developing parallelization techniques for multi-core processors, as well as for IBM’s Cell Broadband Engine. “The newly proposed methods lead to near-ideal speedups on eight core multiprocessors, and to superlinear speedups on the Cell,” says Sandu. “These results are very encouraging, as they allow air quality models to fully benefit from the recent developments in computer chip technology.”
Reality check: Merging information from measurements and models
Significant advancements have been made in our ability to measure key trace gases and aerosols using sensors installed at surface sites on such mobile platforms as boats and airplanes, and on satellites. The European Space Agency estimates that, by the year 2020, every point on and above the Earth’s surface will be viewed from space with a resolution better than 1 kilometer in distance and 1 minute in time.
“The increase in both available data and in the computational power available to process it offers an unprecedented opportunity for a scientific breakthrough in the ability to rapidly integrate measurements and make detailed predictions from complex chemical transport models,” says Sandu. “The close integration of observational data that is recognized as essential in weather forecasting will be as important in improving air quality and climate predictability.”
Data assimilation is a rigorous procedure that seeks to combine model forecasts with observational data streams in order to produce a coherent, accurate sequence of reports of the evolving atmosphere. “The model encapsulates our scientific understanding of many complex, non-linear atmospheric processes and their interactions; the model brings consistency with the physical and chemical laws that govern the evolution of the atmosphere,” says Sandu. “On the other hand, the actual observations are snapshots of reality used to correct the predictions of the imperfect model; the data brings consistency with the real world.
“Although the idea of combining model predictions and observations is common sense, its full realization is extremely challenging, and is the focus of intense research at the cutting edge of scientific computing,” Sandu says.
With support from the National Science Foundation, including a National Science Foundation CAREER award, as well as grants from the National Aeronautics and Space Administration (NASA) and the National Oceanic and Atmospheric Administration (NOAA), Sandu’s group works on developing the computational tools necessary to make chemical data assimilation a day-to-day reality.
“The traditional computer simulations are deterministic and do not quantify the level of confidence associated with their imperfect results. A departure from this traditional paradigm is required in order to rigorously combine models and observations. Model predictions as well as observations are uncertain, and are best described in a probabilistic framework. We know we don’t know everything about the atmosphere, and we seek to mathematically represent the limits of our knowledge,” Sandu emphasizes.
In each data assimilation cycle, a model forecast is produced that represents the time-evolved state of the atmosphere together with its estimated uncertainty. The forecast is then combined with the observations and their errors to produce an analysis, or a best estimate, of the current state and its associated uncertainty.
Considerable challenges are posed by the huge number of degrees of freedom, the highly nonlinear dynamics of the system, and the indirect relationship between the measured quantities and the physically-relevant variables, such as the amount of radiation scattered at different frequencies to obtain information about the column ozone concentrations.
Accurate forecasts depend upon good estimates of the initial conditions, among other critical model parameters. These are determined from measurements through a technique called “inverse modeling.” To make this possible, the Computational Science Laboratory is constructing additional models, called the tangent linear and the adjoint, which are as complex as the original model and whose solutions represent derivatives of the original forecast.
The laboratory is also able to estimate critical model parameters by sampling 50 possible scenarios of the atmospheric evolution. But how can one sample a 10-million-dimensional space with only 50 points and obtain statistically meaningful results?
“Special techniques are being explored for this,” says Sandu. “Careful periodic recalibrations of the ensemble spread prevent the filter from trusting the model too much and ignoring actual information from new observations. The information from each sensor is used to adjust only nearby gridpoints; without this ‘localization’, the predicted ozone over North America may be erroneously increased by an observed high pollution episode in Europe. Our result with this technique is a significant improvement of the analysis and the forecast solutions,” he says.
The CSL methodology to obtain top-down estimates of pollutant emission rates from observations of the concentrations in the atmosphere has important implications. One is to control man-made emissions in order to comply with the U.S. Natural Ambient Air Quality Standards. Another is to better understand the carbon and hydrological cycles in order to improve the accuracy of future climate predictions and to support implementation of the Kyoto Protocol.
In addition to the EPA’s CMAQ model, tools developed by CSL are being used in such large-scale data assimilation studies as Geos-CHEM, a Harvard/NASA model that is used worldwide as a framework for global chemical transport investigations.
As many cities in the U.S. are implementing warning systems for high-pollution episodes, the benefits of chemical data assimilation will be felt daily through improved “chemical weather” forecasts. A deeper understanding of the global impacts of man-made pollution, such as the Antarctic ozone hole, will lead to better environmental policy decisions.
The Earth viewed from the outside
The design of instruments to be deployed onto future satellites depends on the answer to the critical question: What is the added value of the new measurements? A major challenge is to mathematically quantify the information benefit and observation strategies of the proposed instruments well before the satellite is launched.
After the satellite is deployed, the next major challenge is to validate its measurements. No single sensor, instrument, platform, or network can provide all of the information necessary to address this issue. Well orchestrated “sensor webs,” made up of constellations of spacecraft, integrated airborne campaigns, and distributed ground sensor networks, have been actively pursued to achieve the needed multi-dimensional observation coverage.
In collaboration with researchers at the NASA Jet Propulsion Laboratory, Sandu’s group is creating the Sensorweb Operations Explorer (SOX), a virtual framework that can support orbital and sub-orbital observation system simulation experiments.
Virtual experiments with SOX use a complex global chemical transport model to provide high-resolution simulations of the Earth atmosphere. Virtual measurements are taken by simulated instruments on board virtual satellites. This framework allows scientists to rapidly explore the highly complex design space and make the best decisions about instrumentation to be deployed.
“Being able to peek — through model forecasts — at possible future evolutions of the Earth reminds everyone, myself included, to be respectful of the world we live in,” Sandu says.
For more information about the Computational Science Laboratory, please see http://csl.cs.vt.edu.
