The next wave of the Industrial Revolution, Industry 4.0, is underway and relies heavily on real-time data, machine learning, and automation. Technological evolutions help manufacturers improve operational efficiency and reduce wasted time and valuable resources.
In January, scientists at Argonne National Laboratory in Illinois began working with Columbus, Ohio-based industrial tech organization EWI, GE Research, and GKN Aerospace on a two-year U.S. Department of Energy-funded project to develop a better way to manufacture airplanes by leveraging machine learning techniques. Their goal is to reduce the number of pricey experiments and simulations, thus saving time and resources.
“Each one of us are coming up with a specific and unique set of tools to improve aerostructure manufacturing,” Prasanna Balaprakash, Argonne computational scientist, told Centered. “Our national lab has a specific set of tools and technologies — in particular, machine learning and supercomputing capabilities.”
The research focuses on manufacturing aircraft parts with aluminum-lithium alloys, which are attractive for aerostructure manufacturing because of their excellent stiffness and strength. That helps a plane withstand daily rigors such as turbulence. Yet the alloys also are lightweight, which helps to reduce fuel consumption.
Each aircraft component is manufactured individually and then welded to other components to create the full-scale product. Manufacturers have to optimize the production and welding processes for each part. But the optimization process itself exhibits inefficiencies.
Operators try different settings on the manufacturing equipment for each component and examine how the tweaks affect qualities such as stiffness and strength.
“It’s a very trial-and-error process which requires not only time but also resources,” Balaprakash said. “If you do something wrong, the part that you tried to manufacture may not be used.”
The team is working to develop an automated process to reduce the trial-and-error nature of machine optimization for each component. They aim to use machine learning — the process a computer uses to train itself to find the best solutions to a particular question — so the system quickly reaches conclusions about optimal manufacturing settings instead of a human having to program and test all possible settings.
The scientists input real-world aircraft part manufacturing data covering various parameters and conditions to create a predictive model. The computer learns the relationship between data inputs and outputs from the real-world experiments and uses that as a computational model to predict the best manufacturing settings.
“It’s pretty much like how you buy products on Amazon: You buy certain products and based on that you get recommendations,” Balaprakash said. “Or you watch certain movies on Netflix and based on that you get a recommendation.”
The computer can try millions of different combinations in seconds and at a fraction of the cost of traditional optimization.
Balaprakash and his Argonne colleagues have spent about three years developing an open-source machine learning platform called DeepHyper. They’re using DeepHyper and supercomputers to automatically come up with the right predictive model for the airplane manufacturing project, instead of a human laboriously trying to identify the right algorithm.
“Instead of having 100 trials to get it right, you can do it in, let’s say, 10,” Balaprakash said. “We can reduce the energy and resources involved … if we manage to do this with a predictive model.”
The research has the potential to be transformational for manufacturing efficiency. Many industrial processes already incorporate machine learning, but not at the scale of this project.
“What is important is our ability to use the supercomputers at Argonne to do this,” Balaprakash said. “It was not possible before, and these types of resources were not available to the manufacturing industry before.”
The algorithm being developed is specific to aircraft manufacturing. However, the researchers say the machine learning optimization concept can apply to other industrial sectors and manufacturing processes to improve their efficiency and resource conservation.