Ann Arbor Times

Researchers at the University of Michigan have made significant strides in demonstrating that mechanical systems can learn, expanding the realm of machine learning beyond digital platforms. Physicists Shuaifeng Li and Xiaoming Mao have developed an algorithm that allows materials to perform tasks by themselves, such as identifying different species of iris plants.

"We're seeing that materials can learn tasks by themselves and do computation," said Li, a postdoctoral researcher involved in the study. The algorithm is based on backpropagation, a method previously used in digital and optical systems. This new approach could potentially offer insights into how living systems learn.

Li and Mao's research focuses on mechanical neural networks (MNNs), which operate similarly to artificial neural networks but use physical inputs like weights affixed to materials. These materials change shape in response to these inputs, effectively processing information without traditional computing power.

"The force is the input information and the materials itself is like the processor, and the deformation of the materials is the output or response," explained Li.

The researchers used rubbery 3D-printed lattices for their experiments. These lattices are composed of tiny triangles forming larger trapezoids. By adjusting segments within these structures, they demonstrated how MNNs could be trained to respond differently based on input forces.

In one example, they used datasets of input forces related to iris plant features to train their MNNs. The system was then able to correctly identify unknown species based on its training.

Looking ahead, Li aims to increase system complexity using sound waves as inputs due to their ability to encode more data through amplitude, frequency, and phase variations. The team also explores broader network classes in polymers and nanoparticle assemblies with hopes of achieving fully autonomous learning machines.

This research received support from the Office of Naval Research and National Science Foundation Center for Complex Particle Systems (COMPASS).