Steve Koutsourelakis wants to put a number on uncertainty.
Uncertainty in predicting the behavior of structural materials abounds. It’s not just that designers don’t know if a structure will be subjected to an earthquake, they don’t know when in the life span of the structure the earthquake could strike, its magnitude, duration, or frequency spectrum. Core samples can give an indication of the properties of the soil on which a structure is built, but not a complete picture.
“There is uncertainty in the data that you have to design a system and make sure it will be safe,” says Koutsourelakis. “I’m very interested in quantifying these sources of uncertainty and introducing them into the analysis process and quantifying their effects in terms of the probability of failure.”
Uncertainty is crucial when designing structures that must have a small probability of failure—such as nuclear power plants—but not waste millions of dollars. “Buildings in Ithaca have a chance of experiencing a big earthquake, but we don’t build 10-foot-thick walls,” says Koutsourelakis. “You have to find a balance and to do that you need to put a number on these things.”
Getting a better handle on uncertainty could not only help predict how materials will behave, but also provide a clearer picture of how it actually behaves. Sensors can monitor bridges, but uncertainty must be factored into the data to gain an understanding of the overall health of the structure and predict its life span. “Even if you have hundreds of sensors, there’s still uncertainty because you’re not measuring every possible point,” explains Koutsourelakis. “You need to fill in the gaps based on the data you have collected and make predictions about how it will evolve.”
Dealing with uncertainty takes supercomputers, and that’s just to model the behavior of a small sample of material for a brief instant at the atomic level. To expand that model to a structure as big as a skyscraper that must stand for decades is beyond the capability of current processors. But Koutsourelakis thinks an idea borrowed from communication theory might be able to help.
Similar to the way a digital audio file can be converted to an MP3 without losing significant quality, atomistic data might also be compressed to make it more manageable. “The challenge is how do you transfer the information from the fine scale to the scale of interest without compromising the informational content of these fine details and in a manner that allows you to make good predictions about the behavior at coarser scales,” he says.
The chance to work on problems like this in a multi-disciplinary environment is what drew Koutsourelakis to Cornell. “It’s a school that’s extremely good in so many different areas and fields,” he says. “Cornell offers an environment that allows you to interact with so many different disciplines.”
Prof. Koutsourelakis' Web page
