Elastic metamaterials are formed by a spatial arrangement of small-scale unitary elements, yielding unusual dynamic characteristics at the macro-scale. Each element can be seen as a “mirror”, that is capable of refracting or stopping waves and vibrations to the next-neighbours. From a topology viewpoint, artificial neural networks (ANNs) and metamaterials share a similar structure, whereby a number of elements are functionally connected to produce desired input-output relations. However, this task usually requires extensive data processing in digital domain and a relevant amount of computational power.
The goal of this thesis is to develop meta-micro electromechanical structures (metaMEMS) that emulate the behavior of a neural network, thereby perfoming computation without the need of digital architecture and external power. In addition, physical neural networks can be designed to operate “training-free”, i.e. without the need for a back-propagation training algorithm, making this research very attractive for low-power and low-cost intelligent devices.
