AI material that learns behavior and adapts to changing conditions

Just like a pianist who learns to play his instrument without looking at the keys, or a basketball player who invests countless hours to throw a seemingly effortless jump shot, UCLA mechanical engineers have developed a new class of materials that over the course of can learn behaviors over time and develop its own “muscle memory” that allows for real-time adaptation to changing external forces.

The material consists of a structural system of tunable beams that can change shape and behavior in response to dynamic conditions. The research, which has applications in building construction, aircraft and imaging technologies, among others, was published in Wednesday Scientific Robotics.

“This research introduces and demonstrates an artificial intelligent material that can learn to exhibit desired behaviors and properties when exposed to increased environmental conditions,” said Jonathan Hopkins, professor of mechanical and aerospace engineering at UCLA’s Samueli School of Engineering who led the research. “The same basic principles used in machine learning are used to give this material its intelligent and adaptive properties.”

For example, if the material is placed in airplane wings, it could learn to change the shape of the wings based on wind patterns during a flight to achieve greater aircraft efficiency and maneuverability. Building structures infused with this material could also self-adjust stiffness in specific areas to improve their overall stability during an earthquake or other natural or man-made disaster.

Using and adapting concepts from existing artificial neural networks (ANNs), which are the algorithms that power machine learning, researchers developed the mechanical equivalents of ANN components in a networked system. The mechanical neural network (MNN), as the team called it, consists of individually tunable beams aligned in a triangular grid pattern. Each beam features a voice coil, strain gauges, and flexures that allow the beam to change its length, adapt to its changing environment in real time, and interact with other beams in the system.

The voice coil, which takes its name from its original use in loudspeakers to convert magnetic fields into mechanical motion, initiates finely tuned compression or expansion in response to new forces acting on the beam. The strain gauge is responsible for collecting data from the beam’s movement, which is used in the algorithm to drive the learning behavior. The flexures essentially act as flexible joints between the movable beams to connect the system.

An optimization algorithm then governs the entire system, taking the data from each of the strain gauges and determining a combination of stiffness values ​​to control how the network should adapt to applied forces.

To verify the validity of the strain gauge-monitored system, the research team also used cameras trained on the system’s output nodes.

Early prototypes of the system showed a delay between the input of the applied force and the output of the MNN response, affecting the overall performance of the system. The team tested several iterations of the strain gauges and bends in the beams, as well as different grid patterns and thicknesses, before arriving at their published design, which managed to overcome the lag and accurately distribute the applied force in all directions.

“Find out the reasons why [the networks] not learning is important to understand how to design MNNs that learn successfully,” the researchers shared, sharing how they’ve been solving the problem through trial and error over the past five years.

Currently, the system is about the size of a microwave oven, but researchers plan to simplify the MNN design so that thousands of microscale networks can be fabricated within 3D lattices for practical materials applications. Aside from using the material in vehicles and building materials, the researchers suggest that MNNs could also be incorporated into armor to deflect shock waves or into acoustic imaging technologies to harness sound waves.

The paper’s lead author, Ryan Lee, is a graduate student in mechanical and aerospace engineering and a member of Hopkins’ Flexible Research Group at UCLA. Also involved in the research was Erwin Mulder from the University of Twente in Enschede, the Netherlands.