Neuromorphic computing using spiking neural networks (SNNs) offers energy-efficient computing paradigms. Hardware implementations of SNNs using memristors have shown promise, but an interface to convert analog sensor signals into SNN spikes (afferent nerve) is lacking. Traditional computing architectures face challenges due to the von Neumann bottleneck, especially with the increasing volume of data from the Internet of Things (IoT). SNNs, with their energy efficiency and computing capacity, present a compelling alternative. While various technologies are being explored for hardware SNNs (digital logic, CMOS analog circuits, and memristors), the crucial role of an afferent nerve remains unaddressed. Memristors are attractive due to their high integration density, low power consumption, analog behavior, and diffusive dynamics. Memristor-based artificial synapses and neurons have been developed, but a bio-inspired afferent nerve to convert analog signals to spikes is needed. While some progress has been made with organic ring oscillators (OROs), a more compact and efficient solution using memristors is desirable. The NbOx memristor, with its negative differential resistance (NDR) behavior, is a strong candidate for this application, as it can support dynamic threshold switching and emulate biological neurons. This research focuses on creating an artificial spiking afferent nerve (ASAN) using a specially designed NbOx memristor. The ASAN's key function is to translate analog input signals into correlated spiking frequencies, mimicking the behavior of biological afferent nerves. The spiking frequency of this ASAN is designed to be proportional to stimuli intensity under normal conditions but decrease under excessively strong stimuli, thus reflecting a protective mechanism found in biological neurons. A power-free spiking mechanoreceptor system using a piezoelectric sensor is also developed to showcase the ASAN's capabilities in a practical application.

The paper reviews existing neuromorphic computing architectures and their limitations. It highlights the von Neumann bottleneck and the advantages of SNNs in energy efficiency. The literature review emphasizes the use of memristors in SNN implementations and their advantages over traditional transistors. Previous work on artificial synapses and neurons based on memristors is also summarized. The authors discuss the need for an afferent nerve interface to connect SNNs with the environment, referencing existing approaches using organic ring oscillators and other devices for emulating nociceptors and mechanoreceptors. The NbOx memristor's properties and its use in previous neuromorphic computing applications are presented, setting the stage for the authors' proposed ASAN based on this technology.

The research involved the fabrication of a 3D vertical metal-insulator-metal (MIM) NbOx memristor device with a low thermal conductivity polysilicon (poly-Si) bottom electrode. The poly-Si electrode was designed to reduce the threshold current. Transmission electron microscopy (TEM) and energy dispersive spectroscopy (EDS) were used to characterize the device structure and composition. A DC sweep was used to precondition the device, and its switching behavior was analyzed under triangular voltage sweeps. A compact ASAN circuit was constructed using the NbOx memristor and a resistor (Rc). The ASAN's functionality was tested using various input signals (square, triangular, and sinusoidal pulses) to examine the relationship between input voltage and output spiking frequency. The ASAN's energy consumption per spike was calculated. An external capacitor was added to the ASAN to simulate the integration of analog signals, and the system's response to these inputs was analyzed. A power-free artificial spiking mechanoreceptor system (ASMS) was developed using the ASAN and a piezoelectric device as the tactile sensor. The system's response to various pressure levels was experimentally evaluated. A biphasic memristor model was implemented in LTspice for simulation and compared to experimental results.

The researchers successfully fabricated a 3D NbOx memristor with a low threshold current (~2 µA) due to the poly-Si bottom electrode. TEM analysis confirmed the formation of a crystalline NbO2 channel during the forming process. The ASAN, composed of the NbOx memristor and a resistor, demonstrated an oscillation behavior with a spiking frequency proportional to the input voltage up to a certain point. Beyond this threshold, the spiking frequency decreased, mirroring the protective inhibition mechanism observed in biological neurons. The ASAN exhibited quasi-linear frequency response to square, triangular, and sinusoidal input pulses. The minimum energy consumption per spike was approximately 38 pJ. Using an external capacitor, the ASAN demonstrated successful operation with analog input signals, showing frequency increases with increasing voltage up to a point where the frequency decreased before the oscillation ceased. The ASMS, using the ASAN and a piezoelectric sensor, functioned without an external power source, translating pressure intensity into spiking frequency. The ASMS also exhibited protective inhibition under high pressure. The results demonstrated the successful emulation of biological afferent nerve behavior in a compact and energy-efficient memristor-based ASAN. The simulations using a biphasic memristor model closely matched the experimental results. The endurance test showed the ASAN could function for >10¹² cycles. A wider frequency range (0 Hz to 1100 Hz) was also achieved by changing the external capacitor to 47nF, which matched the human nervous system.

The successful creation and testing of the ASAN and ASMS address the need for a robust and efficient interface between sensors and spiking neural networks. The ASAN's ability to mimic the behavior of biological afferent nerves, including protective inhibition, is significant. The power-free operation of the ASMS shows promise for low-power applications, particularly in robotics. The close correlation between experimental results and simulations validates the memristor model and suggests the feasibility of using the ASAN in more complex systems. The energy efficiency demonstrated by the ASAN (38 pJ per spike) makes it highly suitable for large-scale integration and applications requiring minimal power consumption. The ability of the ASAN to work with various types of input signals suggests its versatility in different sensing applications. This research makes a significant step towards creating self-aware neurorobotics and building more sophisticated neuromorphic systems.

This work successfully demonstrated a novel artificial spiking afferent nerve (ASAN) using NbOx Mott memristors. The ASAN emulates the behavior of its biological counterpart, accurately converting analog signals into spiking frequencies while exhibiting protective inhibition. The integration of the ASAN into a power-free mechanoreceptor system proves its practical applicability. Future research could focus on integrating the ASAN with other types of sensors and incorporating it into larger-scale SNN architectures for more complex tasks. Exploring different memristor materials and optimizing the device design could further improve energy efficiency and performance.

The current study focuses on a specific type of memristor and sensor. The generalizability of the findings to other memristor technologies and sensor modalities requires further investigation. The ASMS's response might be affected by the piezoelectric device's characteristics and long-term stability. While the endurance was tested, long-term reliability under continuous operation needs further study. The current ASAN design may require further optimization for integration into large-scale SNNs. More research is needed to evaluate its performance and robustness in real-world applications.