A meso-scale, thermally-actuated mechanical neural network
Intelligence is defined as the ability of a given system to 'find the connection between various things'.
This can refer to the human aptness at figuring out mathematical problems of varying complexity, the way bees communicate among each other through their movements, a neural network's ability to spot golden retrievers in pictures, and many other things.
​
The principles behind this phenomenon (interconnectedness and tunability) are the foundation for mechanical neural networks, intelligent mechanisms capable solving mechanical problems.
Given an arbitrary set of desired mechanical behaviors (a combination of forces and displacements on a given structure), this new type of mechanism, is capable of finding the best solution given a set of constraints (the connections, size, geometry of the beams in a given structure) by varying some of its individual internal properties (the stiffness of each beam).
​
Applications of this new type of mechanism include plane wings capable of adapting to varying air conditions, buildings that can resist earthquakes, and other shape-morphing structures and metamaterials.
​
My current work with regards to mechanical neural networks pertains to:
-
Studying optimization algorithms
-
Designing smaller and more compact versions of mechanical neural networks
-
The ability of mechanical neural networks to be used as a design tool for non-homogeneous metamaterials
-
The ability of mechanical neural networks at performing generalized learning tasks
​
​​