Adaptive Control of Toxic Substances
By Kim Roth
 Now, when terrorist attacks are a greater threat than ever before, adaptive controls provide an effective way to protect public venues from increasingly dangerous toxins, such as airborne chemicals that can kill in a matter of seconds. |
What if there were a poison gas attack in a mass transportation terminal, shopping mall or other public place—an attack somewhat like the release of sarin, an extremely toxic gas, in Tokyo’s subway system in 1995? Or what if there were a release of chemicals into a municipal water supply? Right now, response times in these scenarios rely largely on manual detection and action and, for that reason, take longer to react to than experts would like.
Nik Katopodes, professor and chair, Civil and Environmental Engineering, is working on an adaptive control method that could handle these sorts of problems swiftly and effectively.
He pointed out that control systems are everywhere—in the clock that wakes you each morning, in the toaster on your kitchen counter, and in the fuel injection, anti-lock brakes and cruise control on your car. Mechanical systems such as these are often designed with the control in mind. They’re typically linear, with the action and reaction proportional and, therefore, highly predictable. They can, in short, be controlled precisely.
Katopodes started his work on the adaptive control in the early 1990s as part of a National Science Foundation project. His objective was to develop an adaptive control algorithm for fluid-related problems, such as flooding or the introduction of unavoidable (i.e., accidental or intentional) hazardous chemicals into rivers and estuaries. It wasn’t an easy task.
Fluids are hard to retrieve after release. Once spilled, Katopodes explained, they have a natural tendency to spread because only their momentum, boundary conditions and energy dissipation govern motion. Control in these cases is very difficult to achieve.
Katopodes wondered what would happen if, along the boundaries where the flow would take place, he installed a sensor that could detect—and a control device that could arrest and reverse—the flow according to an algorithm until it slowed to a normal, safe level. He imagined that such a control wouldn’t rely on human intervention and would react within seconds.
Modeling an Event
For years, engineers have developed forensic models to describe events such as a flood or water contamination. “If I was told about it the next day, I could put together a nicely developed fully three-dimensional, time-dependent computational grid of the whole thing, and simulate it within a millimeter,” said Katopodes, who with others worldwide has conducted such analyses for 20 years.
Given a hypothetical release of toxic gas, a computer program could use a model to run scenarios, randomly changing variables in the domain of the plume until by trial and error it found the right constellation of solutions to eliminate it. “That’s a tremendous task,” said Michael Piasecki (PhD CEE ’94), associate professor of Civil and Architectural Engineering at Drexel University and one of the students who worked on the project. “Thousands of parameters need to be estimated, and they’re not independent of one another, so it’s almost impossible to find through trial and error.”
If it were possible, it would likely take a month of calculations, Katopodes said. “That’s not what we mean by ‘real-time’ control.”
Katopodes and his students set out to develop an algorithm that wouldn’t rely on multiple scenario calculations. Their algorithm would work for hazardous materials in a fluid or in the air—a chemical attack in a subway tunnel, for example—because the physical phenomena behind the spread of airborne plumes is similar to the spread of those in water.
 September 11 has created an unprecedented task for security systems. Katopodes' work is making it possible to transfer a great deal of this burden from human hands to ever-vigilant computerized adaptive controls. |
The team members used the adjoint equation method, developed by Russian mathematician Gurii Ivanovich Marchuk for the control of nuclear power plant emissions. They developed a model that, in the same domain as the plume (represented by a grid with potentially millions of points), calculates the interplay of multiple variables and predicts the sensitivity of the concentration at any given point within the domain. (Sensitivity refers to the change in concentration at any of those points due to a corresponding change in a control action—for example, the reversal of airflow in a particular vent within the system.)
Katopodes and his students wrote code from scratch and incorporated some subroutines from off-the-shelf software. They tested the algorithm and found that, other than computational challenges such as memory problems and the need for time-consuming overnight runs, the program worked. In contrast to the trial-and-error method, the adjoint model delivered sensitivities in one simulation within seconds. “It derived from having an idea,” Piasecki said, “and it worked almost on the spot. We got very lucky there.”
They validated the model, which can run on a standard Pentium PC, by testing all of the alternative scenarios independently—this involved thousands of calculations. They finished in the late 1990s. “We did some great things,” Katopodes said. “We succeeded in creating a model for mass transport systems as well as for spills in rivers and estuaries. We extended the concept to prevent floods. Two PhD students wrote papers. It was a very gratifying theoretical study.”
Theoretical, because a critical component was missing: a tool for detection.
The Missing Link
Late last year, Katopodes and the University began a collaboration with Sensicore, an Ann Arbor-based startup that makes solid-state liquid-chemical sensors with t
echnology developed by Richard Brown, professor and interim chair, Electrical Engineering and Computer Science. The University has licensed the technology to the company.
The collaboration produced a silicon chip with high-density electrochemical-sensor arrays that detect chemicals in air or water. Changing the patterning in the membrane that covers each sensor can customize the arrays. This customization, in turn, allows a selective detection of chemicals in the presence of others, Brown explained.
The chip measures about 4.5 millimeters by 5.5 millimeters (smaller than half of a penny), and its arrays can detect chemicals in concentrations of one microliter or less.
Validating the Model with Sensors
The next step in the project is to validate the algorithm by running it with data from the sensors. “A computer model can’t capture everything that’s happening,” Brown explained. “A model is great as long as the assumptions are proper, but you can’t perfectly predict. If you have real-time feedback from sensors that are spread out (in the domain), they can tell you where the edge of the plume is. They can verify the model and adjust in real time to make better predictions.”
Each sensor in the domain uses an embedded wireless communication device to relay data to the computer running the model. Shortly, sensors might include embedded GPS technology, too, so they’ll identify themselves and their locations at the same time they deliver data. In the mass-transit application, once sensors detect a toxic plume, a computer would trigger the tunnel’s ventilation (and alarm) system, a simple mechanical process similar to that of a household thermostat.
Coupling the model with the sensors is “what makes it real, dynamic,” Katopodes said. “Suppose we make a change (in a controlling variable) that didn’t quite work on the first attempt. The sensors continue to monitor (relay data) and in the next two seconds they predict the new position of the plume—you have yet another shot. In microseconds, they sense, control and correct.”
Real-World Application
After validation of the model with sensors, the next phase is to build a scaled physical model of a tunnel, deploy the sensors (the algorithm will dictate the size of the separation), introduce an undesirable chemical and see how quickly the system arrests the toxin’s spread.
Sal Pace, general manager of Sensicore, said that this work with Katopodes’ algorithm is essential because it will validate the effectiveness of the sensors in these types of applications (the sensors are presently used in water monitoring systems). “You can make wonderful technology that does everything it’s supposed to do in a lab,” Pace said, “but unless you deploy it effectively, it’s not going to be useful.” —E
Kimberlee Roth is a freelance writer who has contributed to the Chicago Tribune, the Chronicle of Philanthropy, the Washington Post and the Gale Group E-Commerce Sourcebook (forthcoming).
