Probabilistic Threshold-Switching Memristor for LIF Spiking and RBM Sampling

Abstract

Probabilistic switching in volatile threshold-switching (TS) memristors is often regarded as variability in deterministic circuits, but, when properly calibrated, it can provide a useful probability source for neuromorphic and probabilistic computing. Here, we investigate an Ag/HfO2/Pt/Ti conductive-bridge RAM (CBRAM) device and quantitatively characterize its voltage-dependent switching probability at room temperature within an investigated pulse-amplitude window of 1.6–2.0 V. Structural and chemical analyses confirm the intended device stack, while repeated pulse measurements reveal a stable sigmoid-like switching-probability response, Pswitch (V), within the investigated operating window. The measured response is fitted with a logistic function to establish a device-calibrated probability model. The volatile turn-on and self-reset characteristics are incorporated into a compact leaky-integrate-and-fire (LIF) circuit model, yielding controllable firing dynamics as a function of input amplitude, pulse width, and membrane capacitance. Separately, the calibrated probability function is used as the Bernoulli activation probability for restricted Boltzmann machine (RBM) Gibbs sampling. Under identical training conditions, the device-fitted sigmoid maintains stable learning with a 2.4% accuracy gap relative to the ideal-sigmoid baseline. These results establish an experimentally calibrated measurement-to-model framework that connects volatile TS characteristics to spiking dynamics and probabilistic sampling within the tested operating conditions.