Mechanical Neural Networks: A New Frontier in Material-Based Learning
Research reveals mechanical neural networks can adapt and learn tasks independently, opening possibilities for smart materials optimising themselves in real-time.
A new study from physicists at the University of Michigan suggests that mechanical systems may be just as capable of learning as their digital counterparts. While the concept of machine learning has become synonymous with computer algorithms and vast datasets, the team has shown that materials themselves can not only process information but also learn to perform tasks—a phenomenon that could transform how researchers approach problem-solving in the physical realm.
This approach relies on “mechanical neural networks” (MNNs), a form of lattice-based structure that can be configured to react to inputs in specific, predictable ways. Much like the artificial neural networks behind popular chatbots, MNNs process incoming signals and produce outputs. Instead of relying on silicon chips and code, however, these networks operate through physical transformations in response to applied forces. According to Shuaifeng Li, one of the researchers, “We’re seeing that materials can learn tasks by themselves and do computation.”
Applying Backpropagation Beyond the Digital Domain
Central to these findings is an algorithm based on the well-established backpropagation technique. Backpropagation has long been used to train digital neural networks, but the University of Michigan team has extended its logic to mechanical systems. They demonstrated that MNNs can be guided by the very same mathematical principles that enable digital networks to refine their responses over time.
“We’re seeing the success of backpropagation theory in many physical systems,” said Li. By using backpropagation to adjust the physical structure of their lattices—made of rubbery materials arranged into triangular segments—the team effectively ‘trains’ these MNNs. In practice, this might involve making specific lattice segments stiffer or more flexible to achieve a desired output. For now, the adjustment process remains manual: researchers print revised versions of these lattices to reflect the required changes. Ultimately, the goal is to produce materials that can alter their own structure in real-time, much as living organisms adapt to environmental conditions.
The researchers have even shown that MNNs can handle more complex tasks. By encoding information—such as the dimensions of petals and leaves from various iris species—into mechanical inputs, they demonstrated how these systems could recognise and categorise different plant species. It is a compelling hint at the breadth of challenges that mechanical neural networks might eventually tackle.
Toward Autonomous Problem-Solving Structures
The implications of mechanical neural networks could extend far beyond novelty. If materials can learn and adapt, then researchers might one day design structures that respond autonomously to changing conditions. Aeroplane wings, for example, could self-optimise their shape for varying airflow without human intervention. Complex machinery might repair or recalibrate itself as conditions shift, removing the need for frequent maintenance and improving overall efficiency.
The technology remains at an early stage, with further research needed to refine the materials and methods. Expanding the concept to carry more information, for instance by encoding data in sound waves, is already under investigation. “With sound waves, you have the amplitude, the frequency and the phase that can encode data,” Li noted. Such work could accelerate the complexity and sophistication of MNNs, enabling them to handle intricate tasks previously thought to be the domain of advanced computing systems.
As the University of Michigan team continues to explore polymers, nanoparticle assemblies, and other classes of materials, the prospect of fully autonomous learning machines edges closer. Their latest study, published in Nature Communications, may also contribute to biological research by inspiring new ways to understand how living systems learn. As Li said, “I think this might also help biologists understand how biological neural networks in humans and other species work.”
This convergence of physics, materials science, and machine learning points towards a future where the boundaries between digital and physical intelligence become increasingly blurred, opening doors to applications that could reshape numerous fields in the coming years.
