Pattern recognition with superconducting wirelet neurons

TL;DR

A shunted superconducting wirelet is introduced as an artificial neuron, representing the simplest possible superconducting neuron implementation, enabling straightforward fabrication, electronic control, and high scalability and demonstrating suitability for neuromorphic tasks.

Abstract

Neuromorphic computing aims to reproduce the energy efficiency and adaptability of biological intelligence in hardware. Superconducting devices are an attractive platform due to their ultra-low dissipation and fast switching dynamics. Here we introduce a shunted superconducting wirelet as an artificial neuron, representing the simplest possible superconducting neuron implementation. This minimal design, a single superconducting channel with a resistive shunt, enables straightforward fabrication, electronic control, and high scalability. The neuron exhibits spiking voltage behavior driven by the interplay of resistive switching and relaxation, with key properties such as threshold, firing frequency, and refractory time tunable via applied current, temperature, and shunt resistance. We further show that the resulting temporal voltage signals can be incorporated into a training algorithm to achieve accurate pattern recognition, demonstrating suitability for neuromorphic tasks. Finally, we discuss on-chip training using similar wirelets with gated synaptic weights, establishing a scalable, energy-efficient building block for cryogenic artificial intelligence hardware, integrable with other emergent superconducting technologies.

Figures (7)

• Figure 1: Simulation insight into the resistive state of a shunted superconducting filament.a, The calculated current-voltage characteristics of a shunted nanowire (width $10\xi$, schematic circuit in inset, with $L_k=500L_{GL}$ and $R_{sh}=0.5R_{GL}$; see details in Methods), with four main phases labeled. b, The spiking temporal behavior of the output voltage, for increased applied current (shown on each curve, in units of $j_{GL}$) within phase II. $V(t)$ curves are vertically displaced for visibility. Insets show the snapshot of the superconducting condensate in the filament before and during the voltage spike. c, Temporal behavior of the output voltage within phase III, showing the transformation of spiking behavior into damped oscillations corresponding to the continuous phase-slip state. d, Zoom-in on the phase-slip state from panel c), with visualization of its Cooper-pair density profile.
• Figure 2: Experimental characterization of a spiking superconducting filament. Transport measurement on a 3$\mu$m-wide NbTiN filament shunted by 1$\Omega$ resistor, at $T=8$K. a, The experimental setup. b, The measured spiking behavior of the voltage, for increasing applied current (shown in mA). Curves are vertically displaced for visibility. c, The death of the superconducting neuron, corresponding to Fig. \ref{['fig:fig1']}c,d. The behavior seen in panels b,c has been validated at different temperatures, as shown in the Extended Data Fig. 1.
• Figure 3: Control of spiking frequency.a, Spiking frequency of the superconducting filament from Fig. \ref{['fig:fig2']}, increasing with applied current, for different temperatures. b, Spiking frequency of the same filament, as a function of applied current, at temperature $T=8$ K, but for varied shunt resistance. The spiking behavior has been validated for all used shunt resistances, as shown in Extended Data Fig. 2. c, The full temperature-current phase diagram of the superconducting neuron from Fig. \ref{['fig:fig2']}. Phase II is the operational range of the neuron.
• Figure 4: Proof-of-concept pattern recognition with three superconducting neurons.a, Schematic diagram of the circuit, where synaptic multiplexing was software-based. b, Simulated performance of three identical filaments (width $10\xi$, and shunt resistances 0.5, 0.75, and 1$R_{GL}$) in a neural network: (i) exemplified output voltages of each neuron corresponding to the same train of input current pulses for pixels 1-9. (ii) Output result after training, for 50 different images of digit 4. (iii) The input-output matrix for 50 test images of each digit, with percentages of correct and erroneous predictions shown. c, Three experimental NbTiN filaments as in Fig. \ref{['fig:fig2']} (width $3\mu$m, and shunt resistances 0.3, 1, and 1.7$\Omega$ top to bottom) in a neural network: (i) exemplified output voltages of each neuron corresponding to the same train of input current pulses for pixels 1-9. (ii) Output result after training, for 1000 different test images of digit 4. (iii) The input-output prediction matrix for 1000 test images of each digit, showing 100% accuracy.
• Figure 5: A complex task realized with same three-neuron network.a, Illustrative examples of hand-written number 8 from the MNIST database. b, The temporal voltage output of the three neurons for the input sequence of 420 image pixels of the image a(i). c, Output result after synaptic processing of 1000 test images of digit 8. d, The full input-output matrix on 1000 test images of each digit, demonstrating 92.9% accuracy on average.