[论文解读] Jamming-controlled stochasticity in metal-insulator switching

Nicolò D'Anna, Nareg Ghazikhanian+6 more|arXiv (Cornell University)|Feb 11, 2026

Advanced Memory and Neural Computing被引用 0

一句话总结

本论文使用在位布拉格X射线光子相关谱来研究VO2神经形态设备，显示阈下擦除与阻塞转变将随机切换转变为确定性，从而实现设备内学习。

ABSTRACT

Understanding and controlling phase transitions is a fundamental part of physics and has been central to many technological revolutions, from steam engines to field-effect transistors. At present, there is strong interest in materials with strongly coupled structural and electronic phase transitions, which hold promise for energy-efficient technologies. Utilizing a structural phase transition and controlling its plasticity naturally leads to built-in memory, a key feature for emulating neurons and synapses in neuromorphic technologies. Here, $ extit{operando}$ Bragg X-ray photon correlation spectroscopy is used to study the evolution of the nano-domain distribution at the micron-scale in neuromorphic devices made from the archetypal Mott insulator vanadium dioxide. It is found that after electrical switching, slow nano-domain reconfiguration occurs on timescales of thousands of seconds and that the domains undergo a jamming transition, offering control over switching stochasticity at the micron scale. More precisely, repetitive above-threshold currents plastically drive the system into a jammed/glassy state where switching becomes deterministic, while sub-threshold currents erase the short-term memory contained in the nano-domain distribution, recovering stochastic switching, thus offering a path for in-device learning. The results illustrate the importance of studying the nanoscale physics associated with phase transitions in strongly correlated materials, even for macroscopic devices, and offer guidance for future device operation schemes.

研究动机与目标

• Investigate nanoscale domain evolution during electrical switching in VO2 neuromorphic devices.
• Determine how domain distributions influence filament formation and switching stochasticity.
• Explore memory effects, erasure, and plasticity under various electrical switching regimes.
• Characterize temperature and current-driven dynamics of insulating/metallic domains using XPCS.

提出的方法

• Employ coherent Bragg X-ray diffraction to track nanoscale VO2 domain distributions during micron-scale device operation.
• Record diffraction with synchronized electrical switching to obtain two-time and one-time correlation functions g2(t) and G(t1,t2).
• Fit correlation decays with a stretched exponential |F(t)| and extract relaxation times tau and stretching exponent beta.
• Vary temperature to assess thermally activated domain dynamics and activation energy from Arrhenius fits.
• Apply sub-threshold and above-threshold currents to observe memory erasure and jamming transitions, including repeated cycling to study learning/plasticity.
• Use Ga ion irradiation to locally defect VO2 and guide filament formation as a comparative device.

实验结果

研究问题

• RQ1 How does the nanoscale distribution of insulating and metallic VO2 domains evolve during electrical switching?
• RQ2 How does sub-threshold versus above-threshold current affect memory of initial domain configurations?
• RQ3 Can repetitive switching drive a jamming transition that converts stochastic switching into deterministic filament formation?
• RQ4 What are the characteristic timescales and activation energies governing domain dynamics, and how do they depend on temperature or electrical drive?
• RQ5 Can the device be trained (learn) through plastic changes in domain configurations, and can memory be erased or reset?

主要发现

• Sub-threshold currents erase the initial domain configuration over thousands of seconds, producing erasable short-term memory.
• Above-threshold currents form a metallic filament and immediately erase the initial domain configuration, but the new configuration remains stable for tens of thousands of seconds.
• A jamming transition from liquid-like to glassy domain dynamics occurs on timescales of thousands of seconds after switching, with beta transitioning from 1 to 1.5.
• Repetitive high-current switching drives a transition from stochastic to deterministic filament formation over ~1000 cycles, evidencing learning/plasticity.
• Activation energies for domain dynamics without field are E_a ≈ 230 ± 70 meV, and domain reconfiguration under temperature is thermally activated with Arrhenius behavior.
• The larger the current, the slower the jamming transition, suggesting current-dependent exploration of domain configurations due to Joule heating.