Design of Superconducting Nanowire-Based Neurons and Synapses for Power-Efficient Spiking Neural Networks

Information processing with very low power consumption and innovative computing paradigms are required for the development of modern technologies in which real-time elaboration of data is needed. Natural spiking neural networks are being explored for their speed and energy-efficiency. In these systems, spikes generated by neurons take the information, and synapses act as local memories and connections between neurons, allowing networks to learn and adapt to external stimuli. Superconducting electronics for its intrinsic low energy dissipation is the perfect candidate for building bio-inspired neuronal systems. It has already been proposed a structure that mimics the spiking behavior of the neuron and can be based on two different superconducting devices, which are able to generate low-power pulses: (1) Josephson junctions; and (2) NbN nanowires. The former are widely used for their high operation speed and low power consumption. The latter are typically exploited for single-photon detectors (SNSPDs), but recently are emerging as a platform for new electronics, thanks to their ability to interface with high-impedance environments. This work mainly focuses on the design, optimization, and characterization of the nanowire-based elements necessary for the realization of a spiking neural network. The spiking behavior of nanowire neurons has been demonstrated experimentally, but further work is needed to improve the controllability of their properties. An artificial synapse has not yet been fabricated and tested, but it has been designed, exploiting the presence of kinetic inductance, a particular effect of NbN nanowires (inductive synapse). It is able to reproduce some characteristics of its biological counterpart, like the variable connection strength, but still presents some lacks for the creation of large and versatile networks. A new structure developed to improves the performances of the inductive synapse is here proposed (nTron synapse), introducing the nano-cryotrons (nTron and hTron): nanowire-based comparators with tunable gain, that use the formation of a localized Joule-heated hotspot to modulate the current flow in a superconducting channel. SPICE models of all the exploited superconducting devices were created, starting from experimental data and the existing model of SNSPDs, to facilitate a correct design of the nTron synapse and find limitations of the network. Moreover, fundamental elements of the neurons and nTron synapses like (1) large kinetic inductors, (2) shunted nanowires, and (3) nTrons, were fabricated and tested. Electrical simulations were also performed to study in depth a possible integration of nanowire neurons with Josephson junction neurons. Merging the two technologies could be useful to increase the overall performances of the network, but it generates also some problems, that are here analyzed.