NASA and Argonne use deep neural network to enhance CFD code for hypersonic flight
09 April 2022
The same propulsion technology that goes into rockets has made hypersonic speeds—Mach 5 and beyond, i.e., more than 5 times the speed of sound—possible since the 1950s. However, to make hypersonic flight more common and far less expensive than a rocket launch, engineers and scientists are working on advanced jet engine designs. These new concepts represent an enormous opportunity for commercial flight, space exploration and national defense: Hypersonic aircraft could serve as reusable launch vehicles for spacecraft, for example.
Before any aircraft is built and tested, computer simulations help determine what is possible. Researchers have long used computational fluid dynamics (CFD) to predict, among other things, how an aircraft in flight will interact with the forces around it. CFD is a scientific field devoted to numerically expressing the behavior of fluids such as air and water.
An aircraft capable of breaking the sound barrier brings new levels of complexity to an already computationally intense exercise. Researchers at the US Department of Energy’s (DOE) Argonne National Laboratory and the National Aeronautics and Space Administration (NASA) are pioneering the use of artificial intelligence to streamline CFD simulations and accelerate the development of barrier-breaking aircraft.
The wild ride of supersonic flight produces similarly wild fluid dynamics. As it exceeds the speed of sound, the aircraft generates a shock wave containing air that’s hotter, denser and higher in pressure than the surrounding air. At hypersonic speeds, the air friction created is so strong that it could melt parts of a conventional commercial plane.
CFD simulations must account for major shifts in air, not only around the plane, but also as it moves through the engine and interacts with fuel. Air-breathing jet engines, as they are called, draw in oxygen to burn fuel as they fly. In a conventional plane, fan blades push the air along. But at Mach 3 and up, the movement of the jet itself compresses the air. These aircraft designs, known as scramjets, are key to achieving levels of fuel efficiency that rocket propulsion cannot. But running them in hypersonic flight, it’s been said, is like lighting a match in a hurricane and keeping it lit.
Because the chemistry and turbulence interactions are so complex in these engines, scientists have needed to develop advanced combustion models and CFD codes to accurately and efficiently describe the combustion physics.—Sibendu Som, interim center director of Argonne’s Center for Advanced Propulsion and Power Research
To simulate how combustion behaves within this volatile environment, NASA has a hypersonic CFD code called VULCAN-CFD. The code processes multidimensional flamelet tables, where each flamelet represents a one-dimensional version of a flame. The data tables hold these different snapshots of burning fuel in one massive collection, which requires a large amount of computer memory to process. In a newly published study, Argonne scientists used machine learning techniques to reduce the intensive memory requirements and computational cost associated with simulating supersonic fuel combustion.
Working with NASA gave us the opportunity to integrate our novel developments in a state-of-the-art CFD code, and also to further improve the developments for more efficient design and optimization of hypersonic jets.—Argonne computational scientist Sinan Demir, a study co-author
The flamelet table, generated by Argonne-developed software, was used to train an artificial neural network. In an artificial neural network, which is a subset of machine learning, a computer derives insights from data the way a human brain would. Here, the network used values from the flamelet table to learn shortcuts to “answers” about how combustion behaves in supersonic engine environments.
The partnership has enhanced the capability of our in-house VULCAN-CFD tool by leveraging the research efforts of Argonne, allowing us to analyze fuel combustion characteristics at a much-reduced cost.—Robert Baurle, a research scientist at NASA Langley Research Center
The approach has been validated in previous studies for subsonic applications. The new research applies it to supersonic and hypersonic problems, using the high performance computing resources at Argonne’s Laboratory Computing Resource Center. DOE’s Office of Science and NASA’s Langley Research Center provided funding.
The paper detailing the new neural network framework, entitled “Deep neural network based unsteady flamelet progress variable approach in a supersonic combustor,” was presented in early January at the American Institute of Aeronautics and Astronautics SciTech Forum. In addition to Demir and Som, other authors of the paper include Argonne’s Prithwish Kundu (who is now with Euler Motors) and Cody Nunno (no longer with Argonne), as well as NASA Langley’s Robert Baurle and Tomasz Drozda.
Sinan Demir, Prithwish Kundu, Austin C. Nunno, Sibendu Som, Robert A. Baurle and Tomasz G. Drozda. “Deep neural network based unsteady flamelet progress variable approach in a supersonic combustor,” AIAA 2022-2073. AIAA SCITECH 2022 Forum. January 2022. doi: 10.2514/6.2022-2073