A thorny aerodynamics problem is about to get a Netflix-style “big data” treatment from a Michigan-led team of engineers. They’re developing a better description of turbulence, which could enable radical, more efficient airplane designs and improve prediction in other fields where chaotic flow comes into play – from the human bloodstream to weather forecasting.
“The general impression within the turbulence modeling community is that ideas for improved models have completely stagnated, especially over the past two decades,” said Karthik Duraisamy, an assistant professor of aerospace engineering at U-M. He heads the project, which includes collaborators at Stanford, Iowa State, Boeing and the Silicon Valley firm Pivotal Inc. “We are going to take a completely new approach.”
To maximize fuel economy in terms of lift and thrust, an airplane wing must consistently maintain smooth airflow just at the point where a little more speed would break the air up into disorganized eddies. When this swirling flow – or turbulence – dominates, efficiency takes a nosedive.
Engineers would like to design wings that reliably operate at peak performance on computers before they spend hundreds of millions of dollars building, testing and tweaking prototypes. “These models are not typically very accurate, causing a great deal of frustration among designers,” said Duraisamy.
The major snag is turbulence. To the chagrin of scientists, its chaotic nature has defied an accurate mathematical description. Turbulent flow can be calculated precisely in situations involving low-speed breezes over small bodies, such as insect flight, but accurate simulations of fast-moving flows typical of airplane flight require too much computing power.
“You can go to the largest supercomputer in the world right now – one that can execute a quadrillion instructions a second – and you can run a simulation on that computer for the next 100 years, and you still may be able to solve just a fraction of a flow that’s important to a large commercial airplane such as the Airbus A380,” said Duraisamy.
By building a model from a database of airflow measurements and computations, he and his team hope to make predictions based on more realistic approximations.
A self-taught model
Duraisamy compares the method to algorithms for recommending films and products. “Netflix has this database with a large number of users who made a large number of choices. Based on this, what is the next choice you are going to make?” he said.
This is known as “machine learning” because humans aren’t developing the prediction software directly – instead, they write an algorithm that can condense information in a database into the predictive model. The accuracy of the model improves automatically with the amount and quality of the data.
“Applying learning techniques in turbulence is not as simple because Netflix- and Amazon-type predictions do not have to obey the laws of physics. So that’s the big challenge – to see how we can build physics into this machine-learning approach,” said Duraisamy.
One way that the team will include physics is by solving the turbulence problem for the simple scenarios that can be calculated without approximations. These solutions – from research groups all over the world – will populate the database wherever possible. For more complex situations, they plan to fill in the gaps with the most reliable experimental measurements they can find. This database will then be used to help provide approximations while solving turbulence model equations.
Some of this data could come from U-M labs, such as those of Luis Bernal, James Driscoll and Mirko Gamba, professors of aerospace. Their teams specialize in taking measurements of flow patterns, using high-speed laser imaging, that can be used to build aerodynamic simulations.
Putting it to work
Initially, the model built by Duraisamy and his colleagues will be applied to flows over airplane wings, but the same database could eventually feed models for other applications in aerodynamics. Also, it should be easy to keep the model up-to-date by re-running the machine-learning algorithms whenever new or improved measurements are added to the database.
If successful, the method could be expanded to other areas, including combustion and applications outside aviation. For instance, designers of many drug delivery systems need to model the pulsing of blood through arteries, which creates turbulence that affects whether drug carriers can bind to the vessel wall. Meteorologists need accurate models for turbulence in the atmosphere as it influences the formation and evolution of storm systems.
This work is funded as part of the first phase of a $1.6 million grant from the Leading Edge Aeronautics Research for NASA (LEARN) program. This collaborative project includes Juan Alonso, Brendan Tracey (Stanford University), Paul Durbin (Iowa State), Philippe Spalart (Boeing) and Pivotal Inc.