&arrow; physics 15, 184
Materials that learn to change shape in response to external stimuli are one step closer to reality thanks to a prototype system developed by UCLA engineers.
Living things are constantly learning to adapt their behavior to their environment in order to thrive regardless of their environment. Except in science fiction movies, inanimate materials are not usually studied. Now, a team led by Jonathan Hopkins of the University of California, Los Angeles (UCLA) has developed what they call a learnable architectural material. . Consisting of a network of beam-like components, the material learns to adjust its structure and form a specific shape under the influence of stimuli. The team said the material could serve as a model system for future “smart” manufacturing.
The material developed by Hopkins and his colleagues is a mechanical neural network (MNN). Scientists believe that if these smart materials can be produced on a commercial scale, they could revolutionize manufacturing, from building to design. For example, an airplane wing made with MNNs can learn to change shape in response to changing wind conditions to maintain the aircraft’s flight efficiency; A house made of MNN can adjust its structure to maintain the integrity of the building during an earthquake; and knit sweaters from MNN can be customized to fit any size person.
For the demonstration, Hopkins and his colleagues built a 2D triangular mesh MNN the size of a microwave oven. The system consisted of 21 beam-like components, each containing a motor and two force sensors (one at the end of the beam). These sensors transmit the deformation information of each beam to a nearby computer. These data are fed into an algorithm to calculate the local stiffness changes required to produce the desired properties and behavior in the material. This information is then sent back to the algorithm, which repeats the process as necessary. A property or behavior is said to be “learned” if the MNN can achieve that property or behavior without external guidance after removing the material from the computer; MNN stores the learned behavior in its architecture.
To test the system, Hopkins and his colleagues forced it to act as if it were being pushed from above or from the side, then shaped it to form a specific 2D outline. They showed that MNNs learned to simultaneously acquire multiple shapes and to maintain those shapes under different loads. They also showed that the MNN was able to learn two other behaviors (lean to the left and lean to the right) and respond to different inputs, suggesting a mechanical “muscle memory.”
One of the study’s participants, Ryan Lee of UCLA, said their demonstration shows that it is possible to create smart materials that take different shapes when subjected to different stresses. He thinks such materials could have a wide range of applications. Lee’s favorite example is the creation of a self-healing spacecraft that can change its structure to repair damage caused by space debris. “If the spacecraft is made of MNN material, it can transform without human intervention,” he said.
Andrea Liu, a physicist at the University of Pennsylvania, agrees that Hopkins, Lee and their colleagues have demonstrated a system that can learn. But he notes that the material still has a way to go before it can be called truly intelligent. The learning step should be computer-independent, he said. Lee agreed, saying the team is discussing how to embed the learning algorithm directly into the sensor. The team also plans to produce small parts for MNN and capture 2D to 3D. Li said such a structure could contain more nodes, which could produce MNNs with greater learning and adaptation capabilities.
Anna Napolitano is a freelance science journalist based in London, UK.
- RH Lee etc.“Mechanical Neural Networks: Architectural Materials for Behavioral Learning,” Science. Robot. 7 (2022).