Like a pianist who can play skillfully without looking at the keys, mechanical engineers at UCLA have designed a new material that learns behavior over time and develops its own “muscle memory,” allowing real-time adaptation to change external forces. The material consists of a structural system with adjustable beams that can change their shape and behavior based on dynamic conditions. The study was published in Science Robotics on the 19th.
Jonathan Hopkins, a professor of mechanical and aerospace engineering at the UCLA School of Engineering who led the research, said the artificially intelligent material can learn about the behaviors and characteristics it should exhibit when exposed to environmental conditions. For example, when the material is placed in an aircraft wing, it can learn the patterns of the wind during flight and change the shape of its own wing to improve the efficiency and maneuverability of the aircraft, while a building structure infused with the material can also self-adjust the stiffness of certain areas to improve its overall stability during earthquakes or other natural or man-made disasters.
Scientists have utilized and adapted existing concepts of artificial neural networks. Artificial neural networks are the very algorithms that drive machine learning, and the researchers developed mechanical equivalents of artificial neural network components in an interconnected system. This mechanical neural network consists of individually tunable beams oriented in a triangular lattice pattern. Each beam has a voice coil, strain gauges and flexures that allow the beam to change its length, adapt to the changing environment in real-time and interact with the other beams in the system.
The optimization algorithm then controls the entire system by obtaining data from each strain gauge and determining the combination of stiffness values. To check the effectiveness of the strain gauge monitoring system, the research team also used a camera trained on the output nodes of the system.
An early prototype of the system exhibited a lag between the input of the applied force and the output of the mechanical neural network response, which affected the overall performance of the system. The team tested multiple iterations of strain gauges and bending in the beam as well as different lattice patterns and thicknesses, and the final design solution overcame the lag and accurately distributed the applied force in all directions.
The system is currently about the size of a microwave oven, but the researchers plan to simplify the mechanical neural network design so that thousands of networks can be fabricated at the microscale within the 3D lattice for practical materials applications.
