New Class of AI Material can Develop a Muscle Memory of its Own

UCLA’s latest AI research learns behaviors and adjusts to changing circumstances and acquire muscle memory.

A new class of material created by UCLA engineers may learn behaviors over time and develop its own “muscle memory,” enabling real-time response to shift external stimuli. The material is made up of a structural system with tunable beams that can change how it behaves and how it takes on certain shapes. According to UCLA Samueli School of Engineering mechanical and aerospace engineering professor Jonathan Hopkins, “this research introduces and demonstrates an artificial intelligent material that can learn to exhibit the desired behaviors and properties upon increased exposure to ambient conditions.” “This material’s smart and adaptive properties are provided by the same fundamental principles that underlie machine learning.”

For improved efficiency and maneuverability, the material might be used in aircraft wings to learn to change the shape of the wings based on wind patterns while the aircraft is in flight. Structures impregnated with this substance could also adapt to natural disasters like earthquakes. The group used ideas from existing artificial neural networks (ANNs), which are the algorithms that power machine learning, and modified them to create mechanical representations of ANN elements in a networked system. A triangular lattice structure of individually programmable beams makes up the mechanical neural network (MNN). A voice coil, strain gauges, and flexures are built into each beam, allowing it to modify its length, instantly adjust to its environment, and communicate with other beams in the system.

In response to fresh forces applied to the beam, the voice coil starts the fine-tuned compression or expansion. The strain gauge is in charge of gathering information from the motion of the beam that is employed in the algorithm to govern the learning behavior. After extracting information from each strain gauge and combining it with other stiffness variables to determine how the network should respond to applied forces, an optimization algorithm controls the overall system.

Early system prototypes showed a delay between the applied force’s input and the MNN response’s output, which had an impact on the system’s overall performance. Before arriving at the published design that fixed the issue, the researchers tested numerous iterations of the strain gauges, flexures in the beams, and various lattice patterns/thicknesses. The team wants to reduce the complexity of the MNN design so that thousands of networks may be produced on the micro-scale within 3D lattices for useful material applications. Currently, the system is about the size of a microwave oven. MNNs may also be used in armor to absorb shockwaves and in acoustic imaging techniques to capture soundwaves.

The study’s results, which could have an impact on aviation, imaging technologies, and building construction, were released in Science Robotics on Wednesday. The experimental investigation, according to the authors, lays the foundation for AI-architected materials that can be applied to the creation of structures, airplanes, and imaging technologies. The first theoretical concept for neural networks was made in 1944 by Warren McCullough and Walter Pitts, two University of Chicago researchers who ultimately transferred to MIT in 1952. Numerous simple processing nodes that are intricately connected and loosely fashioned after the human brain make up a neural network, which can consist of hundreds or even millions of them. Most current neural networks consist of node layers with an input layer, one or more hidden layers, and an output layer.

Before this study, previous researchers proposed acoustic metamaterials like the acoustic analog computing (AAC) system, but as they are not neural networks, they are unable to learn. An acoustic metamaterial that imitates the behavior of a trained neural network was proposed by Tyler Hughes et al. in 2019. However, while training is done during the design phase by modeling the modification of the mass within a vibrating plate, a manufactured version of the suggested design was unable to learn new behaviors. This is the first time the mechanical neural network concept outlined in this UCLA paper has been physically and experimentally proven, despite the fact that alternative mechanical approaches have also been proposed and tested during the last two years using just simulation.

Share This Article

Do the sharing thingy

About Author

More info about author

UCLA’s latest AI research learns behaviors and adjusts to changing circumstances and acquire muscle memory.

A new class of material created by UCLA engineers may learn behaviors over time and develop its own “muscle memory,” enabling real-time response to shift external stimuli. The material is made up of a structural system with tunable beams that can change how it behaves and how it takes on certain shapes. According to UCLA Samueli School of Engineering mechanical and aerospace engineering professor Jonathan Hopkins, “this research introduces and demonstrates an artificial intelligent material that can learn to exhibit the desired behaviors and properties upon increased exposure to ambient conditions.” “This material’s smart and adaptive properties are provided by the same fundamental principles that underlie machine learning.”

For improved efficiency and maneuverability, the material might be used in aircraft wings to learn to change the shape of the wings based on wind patterns while the aircraft is in flight. Structures impregnated with this substance could also adapt to natural disasters like earthquakes. The group used ideas from existing artificial neural networks (ANNs), which are the algorithms that power machine learning, and modified them to create mechanical representations of ANN elements in a networked system. A triangular lattice structure of individually programmable beams makes up the mechanical neural network (MNN). A voice coil, strain gauges, and flexures are built into each beam, allowing it to modify its length, instantly adjust to its environment, and communicate with other beams in the system.

In response to fresh forces applied to the beam, the voice coil starts the fine-tuned compression or expansion. The strain gauge is in charge of gathering information from the motion of the beam that is employed in the algorithm to govern the learning behavior. After extracting information from each strain gauge and combining it with other stiffness variables to determine how the network should respond to applied forces, an optimization algorithm controls the overall system.

Early system prototypes showed a delay between the applied force’s input and the MNN response’s output, which had an impact on the system’s overall performance. Before arriving at the published design that fixed the issue, the researchers tested numerous iterations of the strain gauges, flexures in the beams, and various lattice patterns/thicknesses. The team wants to reduce the complexity of the MNN design so that thousands of networks may be produced on the micro-scale within 3D lattices for useful material applications. Currently, the system is about the size of a microwave oven. MNNs may also be used in armor to absorb shockwaves and in acoustic imaging techniques to capture soundwaves.

The study’s results, which could have an impact on aviation, imaging technologies, and building construction, were released in Science Robotics on Wednesday. The experimental investigation, according to the authors, lays the foundation for AI-architected materials that can be applied to the creation of structures, airplanes, and imaging technologies. The first theoretical concept for neural networks was made in 1944 by Warren McCullough and Walter Pitts, two University of Chicago researchers who ultimately transferred to MIT in 1952. Numerous simple processing nodes that are intricately connected and loosely fashioned after the human brain make up a neural network, which can consist of hundreds or even millions of them. Most current neural networks consist of node layers with an input layer, one or more hidden layers, and an output layer.

Before this study, previous researchers proposed acoustic metamaterials like the acoustic analog computing (AAC) system, but as they are not neural networks, they are unable to learn. An acoustic metamaterial that imitates the behavior of a trained neural network was proposed by Tyler Hughes et al. in 2019. However, while training is done during the design phase by modeling the modification of the mass within a vibrating plate, a manufactured version of the suggested design was unable to learn new behaviors. This is the first time the mechanical neural network concept outlined in this UCLA paper has been physically and experimentally proven, despite the fact that alternative mechanical approaches have also been proposed and tested during the last two years using just simulation.

Share This Article

Do the sharing thingy

About Author

More info about author