Mechanical neural networks (MNNs) are architected materials that learn behaviors by tuning the stiffness of their beams, in the same way that artificial neural networks tune their weights. In collaboration with Prof. Jonathan Hopkins at UCLA, the NDML is scaling MNNs down from tabletop prototypes to microfabricated lattices. We have built meso-scale MNN lattices in which each beam contains MEMS thermal actuators and piezoresistive sensors under closed-loop control, allowing each beam to reach a programmed positive or negative stiffness within hundreds of milliseconds. Current work uses electrostatic comb-drive actuators and co-designed control electronics to push MNN response speeds toward the limits set by the material itself, for applications such as ultra-precision motion stages and adaptive structures.
The figure above shows how an MNN maps onto an artificial neural network, along with a comb-drive MNN lattice design.