Weidong Zhang

Spike-based computing has attracted increasing attention for efficient information processing, creating demand for neuromorphic hardware capable of generating and processing spikes. Ovonic threshold switching (OTS) devices are promising spiking neurons because their threshold-switching behavior and negative differential resistance enable compact spike-generation circuits, while their scalability supports dense integration. However, a circuit-level design framework linking defect dynamics with conduction behavior remains lacking. Here, we experimentally demonstrate an intensity-tuned artificial neuron based on a GeAsTe OTS device integrated with a resistor-capacitor (RC) circuit. The neuron shows a non-monotonic excitatory–self-inhibitory response with applied voltage, phenomenologically analogous to excitatory–inhibitory behavior in biological neurons. The spiking frequency initially increases and then decreases beyond an inflection voltage point. This transition arises intrinsically from competition between RC charging dynamics and conductive filament retention in the OTS device, rather than from a separate inhibitory pathway. The neuron achieves a spiking frequency of up to 1.6 MHz. The behavior is supported by simulations using a dynamical model for diffusive filament-based memristors, accounting for resistance changes from filament formation and rupture. Furthermore, a physics-informed compact SPICE model combining defect-state transitions and filament evolution reproduces OTS switching and predicts neuron dynamics under varying temperatures, internal resistances, and circuit conditions. Finally, circuit simulations of a coupled OTS neuron network demonstrate vertex coloring with phase convergence within approximately 5 μs, highlighting the potential of this framework for compact, energy-efficient brain-inspired computing.

In this work, the effects of high drain current and high drain voltage on the electrical performance degradation and burnout mechanism of gallium nitride (GaN) RF high-electron mobility transistors (HEMTs) under high-temperature operation stress have been investigated in-depth. It was found that gate leakage increases with stress drain current, while the degradation of the saturation drain current and RF output power is influenced by stress drain voltage. The photon emission and scanning electron microscopy (SEM) analyses show that damage mainly occurs at the edges of central gate fingers under high current. The longitudinal electric field at the gate, caused by positive gate voltage, allows channel electrons to overcome the AlGaN/GaN barrier and pass through the AlGaN layer and Schottky junction. This leads to damage and leakage paths that degrade the Schottky contact and increase gate leakage. The results also show that catastrophic burnout occurs in the gate–drain access region, particularly near the drain electrode under high drain voltage. This is due to new defect formation and electron trapping caused by hot electrons near the drain, creating a strong longitudinal electric field. Once this field reaches a critical level, burnout occurs.

Random telegraph signals (RTS) can serve as a physical entropy source for hardware security, stochastic computing, and artificial intelligence, yet the conventional implementations of RTS face critical limitations: CMOS-based approaches suffer from area inefficiency; emerging memristors offer probabilistic switching but lack controllability; while existing spintronic RTS solutions require external magnetic fields with limited tunability. This work introduces a field-free, electronically modulated RTS methodology using a single magnetic tunnel junction. By exploiting voltage-controlled stochastic magnetization switching, we generate intrinsically random RTS with demonstrated cryptographic-grade randomness. The RTS platform successfully executes key exchange protocols and real-time voice encryption. Crucially, the core parameters of RTS—state occupation ratio and transition time constants—can be dynamically modulated via excitation amplitude adjustment. This controllable RTS also successfully achieved image edge detection without post-processing and played a significant role in suppressing overfitting of neural networks. This voltage-modulated RTS paradigm bridges hardware stochasticity with algorithmic demands, establishing a robust foundation for secure and efficient computing systems.

Power law noise (PLN) has promising applications in healthcare and artificial intelligence (AI). In this paper, we propose a hardware solution for generating PLN with tunable long-term dependency, by combining the magnetic tunnel junction (MTJ) based two-state Markov chain (MC) generation and modulation method. PLNs with specific parameter α can be obtained by merging multiple two-state MCs. The merged PLN, with α ranging from 0.2 to 1.8 and adjustable long-term dependency as represented by the autocorrelation function (ACF), is demonstrated by varying the voltage conditions applied to the MTJs. This method provides a versatile hardware solution for constructing stochastic signal in semiconductor IC chips.

In view of the lower thermal conductivity of β-Ga₂O₃, effective heat dissipation is crucial for maintaining its high output power and reliability. This study presents an active thermal management model for β-Ga₂O₃ Schottky barrier diodes (SBDs) using thermoelectric coolers (TECs), employing both experimental and numerical methods. The heat generated by the SBD can be effectively dissipated to the surroundings through the Peltier effect in thermoelectric (TE) materials. Experimental results show that, with a TEC input current of 6 A, the case temperature remains below 25 ^∘C even at an SBD output power of 25 W, achieving a maximum junction temperature reduction of 74.5 ^∘C compared to when the TEC is off. At a TEC input current of 3 A, the net output power of the SBD reaches 11.8 W, corresponding to an improvement ratio of 36.9%. In simulation, TEC parameters are analyzed to optimize the junction temperature. For every 0.1-mm reduction in the height of TE legs, the junction temperature decreases by 5.3 ^∘C. The Seebeck coefficient significantly impacts the junction temperature through co-optimization of material parameters. This work provides guidance for designing and optimizing the thermal management of β-Ga₂O₃ SBDs with TECs.

Matrices are the foundation of science and engineering. For artificial intelligence (AI) and Internet of Things (IoT) tasks, developing a hardware efficient way to find the eigenvector of stochastic matrix (SM) is urgently in need. In this paper, inspired by the divide-and-conquer strategy, we proposed a new hardware architecture, which uses magnetic tunnel junctions (MTJs) to estimate the eigenvector of an n×n SM where n is the power of 2. This approach reduces the required device amount to log ₂ n by converting the larger SM into 2-state sub-SMs which are further represented by stochastic signals generated by MTJs. The validity of this method has been demonstrated and statistically evaluated. This method provides a novel hardware solution to solve mathematic problems using emerging hardware technologies.

The generation of high quality bipolar random signal for unconventional computing and artificial neural networks remains challenging. In this study, we proposed a new method to generate bipolar random signals using two industrial-viable magnetic tunnel junctions (MTJs) operated solely by electrical operations compatible to the conventional memory program/read circuits. The bipolar random signals demonstrate excellent randomness and this method shows satisfying controllability over the average dwell times (ADTs), with good adjustability in bit generation speed and power consumption. This study provides a new practical hardware solution to the generation of high-quality random signals in the post-Moore era.