California Institute of Technology
Emmet Cleary has broad research interests, although they revolve around fluid flows, especially combustion and chemical transport and kinetics.
What’s more focused is his determination to understand these processes via computing – a resolve he arrived at Down Under.
Cleary, a Department of Energy Computational Science Graduate Fellowship (DOE CSGF) recipient, went to Adelaide, Australia, as a Fulbright scholar to research flameless combustion in coal burners. He used computational fluid dynamics software to simulate the process while another student tested it. “We got similar results between my model and his experiment, so it was exciting,” Cleary says.
He’d done both experimental and computational work as a California Institute of Technology undergraduate, “but the Fulbright is when I actually decided I want to do computational science.” Its broad applications align with his similarly expansive interests, Cleary says.
Now back at Caltech as a mechanical engineering doctoral candidate, Cleary works with advisor Tapio Schneider on mathematical techniques to select simulation parameters that produce the most accurate results and to calculate the inherent uncertainty in those choices. He focuses on climate models, in which many parameters – the values for physical properties influencing the simulated processes – often are unknown, leading researchers to set them manually.
“That begs the question: Did you get the best parameters? Did you pick the right ones? That leads to a lot of uncertainty,” Cleary says. He tests new algorithmic tools to determine the best parameter choices for real-world problems.
In one study, Cleary, Schneider and fellow Caltech researcher Andrew Stuart used an idealized global climate model, then focused on a simple two-parameter scheme for precipitation. They wanted to work back from the model’s results to estimate the parameters that generated them, then calculate the uncertainty inherent in those estimates.
Their method starts with an ensemble of parameter choices, then repeatedly updates the ensembles to narrow their distribution to an optimal number for each parameter.
The researchers then quantify the uncertainty in those predictions. “It comes down to computing a standard deviation” for the estimated parameters, Cleary says. “If we have a really small standard deviation of one parameter, that’s clearly one good choice” on which to base a model. “If it’s a huge standard deviation then there’s a range of parameters that could generate the same data, so there’s really no strong dependence on that parameter.” Cleary will present the team’s method at the Society for Industrial and Applied Mathematics Conference on Computational Science and Engineering in Spokane, Washington, in February 2019.
Cleary returned to combustion research, in a sense, for his 2017 practicum at Sandia National Laboratories’ California location. He worked with Jacqueline Chen, a leading combustion scientist, to extend and test an algorithm designed to cope with asynchronous communication between computer processors. This irregular communication can result from the way computers divide up a problem, such as sections of a simulated combustion chamber, so that individual processors can simultaneously calculate physical properties, cutting the time to reach a solution.
Some mathematical approaches can cope with asynchronous communication when the problems are relatively simple. But more complex, nonlinear equations might be more sensitive to asynchrony, throwing off their solutions’ accuracy.
On his practicum, Cleary tested asynchrony-tolerant methods on a range of one-dimensional fluid problems, including traveling waves, turbulence, shocks and flames. He found the techniques could work on these practical fluid problems and introduced errors at insignificant length scales.
A postdoctoral researcher in Chen’s lab had previously tested the techniques on classical problems, Cleary says, “but I was the first to apply them to real equations that people might solve for engineering applications.” He presented the work at an American Physical Society Division of Fluid Dynamics conference in 2017.
Cleary is unsure when he’ll graduate and what field he’d like to work in after that. “There are too many interesting things” to study, he says, but he’s sure to remain in the sciences and perhaps work at a national laboratory.
Image caption: With the right algorithm, a wide-spread ensemble of model parameters (left figure) can collapse very close to their true values after only five iterations (right figure). Credit: Emmet Cleary.