Flat materials that can morph into three-dimensional shapes have potential applications from medical devices to architecture and more. Programming these shape changes requires complex and time-consuming computations.
Researchers from the Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS) developed a platform using machine learning (ML) to program the transformation of 2D stretchable surfaces into specific 3D shapes.
“While machine learning methods have been classically employed for image recognition and language processing, they’ve also recently emerged as powerful tools to solve mechanics problems,” says Katia Bertoldi, professor of applied mechanics at SEAS and senior author of the study. “In this work we demonstrate that these tools can be extended to study the mechanics of transformable, inflatable systems.”
The team began dividing an inflatable membrane into a 10×10 grid of 100 square pixels that can either be soft or stiff, which can be combined in a variety of configurations, making manual programming extremely difficult.
That’s where ML comes in.
The researchers used finite element simulations to sample this infinite design space. Then neural networks used the sample to learn how the location of soft and stiff pixels controls the deformation of the membrane when it’s pressurized.
“Once the machine learning model was trained, we came up with an arbitrary 3D shape and passed it to the model,” says Antonio Elia Forte, a former postdoctoral fellow at SEAS and first author of the paper. “The neural network then outputs the membrane design and the pressure at which we should inflate such membrane to obtain the desired 3D shape.”
The researchers used the new design method to build and test a device for mechanotherapy that can stimulate tissue around a scar to enhance healing and reduce recovery time.
“This platform has potential to quickly and effectively design patient-specific devices for mechanotherapy and beyond,” Forte says. “Before this research, we didn’t know how to use machine learning to unravel nonlinear mappings in inflatable systems, but it turns out that they are very powerful for these purposes.”
“Machine learning could push the boundaries of currently known design strategies and allow us to design and build fully reconfigurable shape-morphing material,” Forte says.
Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS): https://www.seas.harvard.edu