Cracking the Code for materials that can learn

Los Angeles CA (SPX) Dec 10, 2024
While machine learning often conjures up images of algorithms and digital systems, researchers are demonstrating that the potential for learning extends far beyond the digital realm. A team at the University of Michigan (U-M) has developed a mathematical framework showing how physical materials can "learn" and perform computational tasks without electronic or human intervention.

Shuaifeng Li and Xiaoming Mao, U-M physicists, have introduced an innovative algorithm that leverages mechanical neural networks (MNNs) to train materials. These MNNs, structured as lattices, can adapt their properties to solve problems such as species identification among iris plants. "We're seeing that materials can learn tasks by themselves and do computation," said Li, a postdoctoral researcher at U-M.

The algorithm draws from backpropagation, a concept traditionally associated with digital and optical systems. The researchers believe this approach could inspire advancements in material science and even biological research. "I think this might also help biologists understand how biological neural networks in humans and other species work," Li added.

Physical Systems as Neural Networks

The study highlights how physical systems like rubbery, 3D-printed lattices-composed of tiny triangles forming trapezoids-can mimic the functionality of artificial neural networks. Instead of relying on electronic processors, these materials use their structure to process inputs and produce outputs. For example, applying a weight to the lattice serves as the input, while the deformation it causes acts as the output.

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

While futuristic applications, such as adaptive airplane wings that optimize their shape in real-time, remain a distant goal, this research lays the foundation for practical uses. Currently, the materials require manual adjustments, but advances in smart materials could enable full autonomy in the future.

Training Mechanical Neural Networks

The learning process for MNNs is similar to training digital systems. The researchers use the difference between the actual and desired output, known as the loss function, to guide adjustments. In the study, Li and Mao demonstrated how their algorithm enables lattices to autonomously adjust their structure to produce desired responses.

One experiment focused on creating an asymmetric response in a symmetrical lattice. Using their algorithm, the team achieved this by selectively modifying the lattice's segments. They also trained the materials with large datasets, enabling them to differentiate between various species of iris plants based on petal and leaf dimensions.

Looking ahead, Li is exploring ways to increase the complexity of tasks that MNNs can handle. This includes using sound waves, which can encode richer information through variables such as amplitude, frequency, and phase.

"We can encode so much more information into the input," Li said. "With sound waves, you have the amplitude, the frequency and the phase that can encode data."

The team is also studying broader classes of networks, including polymers and nanoparticle assemblies, aiming to create materials that can autonomously learn and adapt to new challenges.

Future Implications

The potential for materials capable of learning extends beyond engineering and physics. Li and Mao believe their algorithm could help biologists better understand how natural neural networks function in humans and other species.

Their work, supported by the Office of Naval Research and the National Science Foundation's COMPASS Center, represents a significant step toward creating autonomous systems that blend physical and computational capabilities.

Research Report:Training all-mechanical neural networks for task learning through in situ backpropagation