A Distance-Encoding Spiking Sensory Neuron Based on Optimized NbO2 Memristor

Abstract

NbO₂ threshold-switch memristors exhibit S-shaped negative differential resistance (S-NDR) and relaxation dynamics, making them attractive building blocks for compact spiking neurons. However, widely used moving-boundary electro-thermal models (e.g., the Pickett benchmark) are numerically stiff and computationally expensive, limiting circuit-level design-space exploration and neuromorphic sensing simulations. This brief presents a physics-constrained lumped electro-thermal compact model for NbO₂ locally active memristors (LAMs), termed the Optimized NbO₂ LAM. The model replaces the moving boundary with a single lumped temperature state and a smooth phase mapping while preserving key switching invariants ( $V_{th}$ , $V_{hold}$ ) that define the locally active S-NDR window. When embedded in a Pearson–Anson RC oscillator, the proposed model reproduces quasi-static $V$ – $I$ characteristics and relaxation oscillations consistent with the Pickett benchmark, while achieving a median $\sim 3.92\times $ simulation speedup. We demonstrate a distance-encoding spiking sensory neuron in which a light detection and ranging (LiDAR)-inspired input-amplitude proxy monotonically modulates the output spike frequency. A field-programmable gate array (FPGA) register-transfer level (RTL) implementation, verified via Vivado XSim functional simulation, confirms a mean frequency error of 0.90% over the 15-point FPGA validation range relative to a discrete Optimized NbO₂ LAM reference. A PySpice transient cross-check further validates the waveform fidelity of the Optimized NbO₂ LAM against a Python ODE solver. A dynamic distance-tracking experiment additionally provides temporal evidence for spike-based distance perception (RMSE $\approx 6.9{\,}\mathrm {cm}$ ), indicating feasibility for edge spiking neural network (SNN) interfaces.