����������

Maryam Shakiba is studying complex composite materials with machine learning to make stronger and lighter aircraft for the Navy.��

Shakiba, an assistant professor in the Ann and H.J. Smead Department of Aerospace Engineering Sciences, is leading a $750,000 grant from the Office of Naval Research, Aerospace Structures and Materials, to use machine learning techniques to advance composites made with additive manufacturing – more commonly known as 3D printing.

“Additive manufacturing has advanced a lot in the last few years,” Shakiba said. “We can now print complex, fiber-reinforced composite materials. Because we can print more complex patterns, we also need fast computational approaches that can model and predict the response of those materials.”

Navy aircraft technology has generally used metal body panels, but are starting to rely more on composite materials, like passenger jets have for years. Modeling the performance of such materials prior to construction is critical to determining their strength and potential failure points.

Traditionally, this requires finite elements analysis, a tried-and-true method of mathematical modeling. However, the complexity of the method demands major computing resources.

“If you have a material and you change one parameter, a finite elements simulation takes a few days. We need faster models to explore the design space better,” she said.

Shakiba’s work in machine learning is opening new opportunities for that modeling.

“We’ve integrated a convolutional neural network and a graph neural network that increases accuracy and decreases the amount of data you need to put in to get good results. The preliminary results show you can reduce the training data by at least 50 percent,” Shakiba said.

Even with a need for dramatically less data, the work requires supercomputers, like ����������’s Blanca cluster, but the results are spit out in seconds instead of days.

Over the course of the three-year grant, Shakiba and her team, which includes partners at Johns Hopkins University, will advance these machine learning tools with increasingly complex composite patterns. The goal is to combine analysis of materials at both micro- and macro-scale to develop a complete picture of a composite’s response to stress.

“There is a huge interest from the federal government in decreasing the amount of time it takes to design to using a material it in the field,” Shakiba said. “Our method can do that.”