Raphael ThysA neural network forms the basis of many modern artificial intelligence set-ups, and now the concept has been applied to a purely mechanical calculating machine
A mechanical neural network built from beams of variable stiffness
A mechanical neural network composed of beams, motors and sensors can learn to carry out several different tasks, just like its software equivalent, and could lead to aircraft wings that morph during flight to maintain efficiency or minimise turbulence.
The basis of modern AI research is the artificial neural network (ANN), which mimics the structure of the human brain by creating large grids of artificial neurons connected by synapses. Just as the human brain learns new behaviours by strengthening synaptic connections, ANNs learn by adjusting the digital values stored to represent them.
Ryan Lee at the University of California, Los Angeles, and his colleagues have borrowed that concept to create a mechanical neural network in which the strength of connections between neurons is replaced by beams of variable stiffness.
Instead of processing digital data, the mechanical neural network processes forces applied to it, twisting and morphing its shape depending on the stiffness of its beams. For instance, an even force applied across one side of the network can be directed by many beams to immediately create force in a wave shape at the opposite end, while an even force applied upwards could produce the inverse wave.
The team built a network of 21 beams, each 15 centimetres long and arranged in a triangular grid. Every beam is equipped with a small linear motor, which can alter its stiffness, and sensors that measure how far each “neuron”, or beam joint, is out of position. This allows a computer to train the network by tweaking the beam stiffness. Once this is done, the structure requires no external computation and the various beam stiffnesses are locked in.
Read more: Simple mathematical trick could slash AI development time in half
Lee says that an aircraft wing made of a mechanical neural network could automatically morph in response to situations, changing its profile in reaction to higher or lower speeds to maintain efficiency or to prevent undesirable flight characteristics.
“You could do something neat like turbulence interference, where the wing gets hit by something and it locally deforms and morphs to try and keep the energy spread out in a way that the cabin feels nice and smooth,” he says. “Right now, wings are designed to do big motions, flex, distribute that across the wing, and that results in jerkiness in the cabin.”
Future versions of the network could be scaled up to a much larger grid, with each beam possibly miniaturised using advanced 3D printing techniques, says Lee. Once a network was trained, either physically or in a simulation, it could be printed with set beam stiffnesses and require no electronics to function from that point on, he says.