Discovery of Emergent Computation in Metamaterials

Abstract

Technology is evolving at great pace, with new AI models requiring vast amounts of power to train and use. Hardware must also evolve to meet these needs, allowing for more energy efficient computation. This thesis investigates material computing, one possible route to more efficient computing. Specifically, this thesis delves into the use of a metamaterial as a substrate for material computing and investigates how it can be tuned and configured to support computation. Artificial Spin Ice, a nanomagnet based metamaterial, was used as a case study to ground the research and provide a concrete problem to approach. Two key tools made this research possible, the first was the development of flatspin a simulator capable of simulating the time evolution of an artificial spin ice quickly and efficiently. The second was the formulation of a representation of spin ice geometries well suited to manipulation via evolutionary algorithms. These tools allowed for the discovery of a variety of emergent behaviours, aided in tailoring these behaviours towards different computational properties, and even tuned a spin ice to display temporal behaviour mimicking structures found in music. Robustness to the noise and disorder of the real world is key if a material computing substrate is to advance from simulation to a physical implementation. A striking result of this thesis is that when the spin ice geometries were tailored to produce a certain computational property they had very poor robustness, yet when they were evolved with added disorder in the system, highly robust geometries were achieved.