As a new engineer at Honeywell Aerospace in Phoenix in 2011, one of Cristina (Tina) White’s first tasks was to computationally model fan blade designs that reduce noise in a gas turbine aircraft engine.
She focused on transonic flutter: vibration resulting from air moving over the blade at or near the speed of sound. In drastic situations, flutter can cause the fan to self-destruct and the engine to fail. “It’s uncommon, but it’s something that really needs to be avoided,” says White, now a Department of Energy Computational Science Graduate Fellowship (DOE CSGF) recipient.
Using computational fluid dynamics (CFD) software, White modeled about 50 iterations for an experimental design. Engineers “came up with new concepts week after week, and I was the person whose job it was to tell them nothing worked,” she says. “I was the bearer of bad news every week for about six months.”
During a second attempt at a design, White developed her own solution – one she was sure other engineers would reject because it could reduce fuel efficiency. Nonetheless, she modeled her concept as a side project and found it worked “exceptionally well. It was basically a way to fix the issue.”
Not everyone was convinced, however, and “there was sort of a battle between well, do we put this into the system or do we not? It does affect efficiency but much less than expected.” In the end, engineers successfully tested White’s scheme. Her name is on Honeywell’s patent, and the change could be part of later designs.
Repeatedly developing and computationally testing concepts showed White how slow CFD can be. Meanwhile, she became fascinated with machine learning, in which computers analyze known data to predict and identify properties in new information.
Now White is a mechanical engineering doctoral student at Stanford University, where she applies machine learning to CFD problems. The goal: improve the speed and reliability of computationally intensive fluid flow and interaction simulations.
CFD is notorious for consuming computer time, limiting its use for aircraft design and similar purposes. One solution is to create a reduced-order model (ROM), which calculates fewer degrees of freedom or dimensions to simplify the simulation. But ROMs must be tailored to each situation, forcing users to tediously adapt their codes, and lose accuracy when the problem is nonlinear, with fewer predictable processes.
Working with her advisor, Charbel Farhat, and with Daniela Ushizima during a 2018 Lawrence Berkeley National Laboratory practicum, White developed an alternative using a neural network, a machine-learning technique inspired by brain structure. Her technique, called a cluster network, trains on results from previously run simulations to predict new solutions.
White compares the process to an experienced engineer’s intuition: by examining simulations, they often can make judgments about what would happen under new parameters. For example, they may model a wing at two different lengths and estimate an interpolated result for a length that falls in between. But the governing process is nonlinear, with no obvious way to model it, White says. “That kind of guessing is hard for machine learning to do well,” but well-designed neural networks can handle the task.
The cluster network finds a way to decompose the problem and quickly interpolate new results. The approach also doesn’t require adapting an existing code.
“It’s easier to implement, and there are lot of cases where it’s faster and more accurate” than methods that start with the governing fluid dynamics equations, White says. A cluster network may not always be better, but its ease of use makes it a good place to start when formulating a CFD solution, she adds. In one online paper, White, Farhat and Ushizima compared a neural network approach to state-of-the-art approximations and found it was nearly as accurate but ran an order of magnitude faster.
“These kinds of networks and machine learning can be applied to so many fields, and the better we do, the more likely it is that we can solve all kinds of problems,” White says. Her application, for now, is aerospace, but “I’m hoping to continue in machine-learning research after I graduate.”
White has slightly delayed commencement to focus on a more pressing subject: the COVID-19 pandemic. She helps lead an international volunteer team that’s developing a privacy-prioritizing mobile phone app that would alert users when they may have been exposed to the coronavirus.
After that, White hopes to do more neural network research that involves automatically breaking nonlinear problems into parts. It’s “an important piece, I think, of what we call intelligence, and that’s why I care about it.”
Image caption: A neural network architecture search for predicting fluids, showing the structures of nodes that develop and strengthen connections from known data, enabling the system to identify and predict properties in previously unseen data. The networks are inspired by biological brain structure. Credit: Cristina White.