UCLA

In complex systems, control and understanding become intertwined. Following Ilya Prigogine, we define complex systems as having control parameters which mediate transitions between distinct modes of dynamical behavior. From this perspective, determining the nature of control parameters and demonstrating the associated dynamical phase transitions are practically equivalent and fundamental to engaging with complexity.

In the first part of this work, a control parameter is determined for a non-equilibrium electrochemical system by studying a transition in the morphology of structures produced by an electroless deposition reaction. Specifically, changing the size of copper posts used as the substrate for growing metallic silver structures by the reduction of Ag+ from solution under diffusion-limited reaction conditions causes a dynamical phase transition in the crystal growth process. For Cu posts with edge lengths on the order of one micron, local forces promoting anisotropic growth predominate, and the reaction produces interconnected networks of Ag nanowires. As the post size is increased above 10 microns, the local interfacial growth reaction dynamics couple with the macroscopic diffusion field, leading to spatially propagating instabilities in the electrochemical potential which induce periodic branching during crystal growth, producing dendritic deposits. This result is interesting both as an example of control and understanding in a complex system, and as a useful combination of top-down lithography with bottom-up electrochemical self-assembly.

The second part of this work focuses on the technological development of devices fabricated using this non-equilibrium electrochemical process, towards a goal of integrating a complex network as a dynamic functional component in a neuromorphic computing device. Self-assembled networks of silver nanowires were reacted with sulfur to produce interfacial "atomic switches": silver-silver sulfide junctions, which exhibit complex dynamics (e.g. both short- and long-term changes in conductivity) in response to applied voltage signals. Characterization of these atomic switch networks (ASNs) brought out interesting parallels to biological neural networks, including power-law scaling in the statistics of electrical signal propagation and dynamic self-organization of differentiated subnetworks. A reservoir computing (RC) strategy was employed to utilize measurements of electrical signals dynamically generated in ASNs to perform time-series memory and manipulation tasks including a parity test and arbitrary waveform generation. These results represent the useful integration of a complex network into a dynamic physical RC device.