Modern electronic devices underpin nearly every aspect of daily life, with their functions recently expanding from simple data storage to acting as artificial synapses in neuromorphic systems. As electronic devices scale down in size, low-frequency no...
Modern electronic devices underpin nearly every aspect of daily life, with their functions recently expanding from simple data storage to acting as artificial synapses in neuromorphic systems. As electronic devices scale down in size, low-frequency noise (LFN) has emerged as a critical issue limiting both performance and reliability. Although LFN has traditionally been viewed as detrimental, this study re-examines this perspective and demonstrates how intrinsic noise in hafnium oxide-based ferroelectric memories can serve as (i) a powerful tool to understand device physics, (ii) a means to optimize device performance, and (iii) a valuable resource for enabling novel computational functions.
Ferroelectric tunnel junctions (FTJs) and ferroelectric field-effect transistors (FeFETs) fabricated with HfO2 thin films were systematically analyzed using LFN spectroscopy. Power spectral density analysis enabled clear differentiation among Poole–Frenkel emission, interface-trap–assisted tunneling, and barrier-height fluctuation mechanisms. Furthermore, the same noise-based approach allowed quantitative discrimination between ferroelectric resistive switching (FE-RS) and non-ferroelectric resistive switching (Non-FE-RS) within a single device. Our findings reveal that the tunneling electroresistance ratio (TER) in FTJs is governed not only by remnant polarization but also significantly influenced by interface and bulk trap states. This insight provides concrete guidelines for optimizing process conditions and programming voltages.
Building on these insights, we propose an adaptive read-bias (ARB) strategy designed to minimize the influence of intrinsic device noise on neuromorphic system performance rather than suppressing the noise itself. By tailoring read voltages according to layer-specific noise sensitivity, ARB reduces power consumption by 61%, increases integration density by 92%, and maintains accuracy degradation below 3%, all without requiring additional correction circuitry.
Finally, we demonstrate a voltage-controlled noise injection layer (FTJ-NIL) that converts intrinsic device LFN into a hardware-level data augmentation engine. Applied to real-world datasets, including medical signals, the FTJ-NIL approach delivers up to a 13-percentage-point improvement in classification accuracy.
The noise-focused characterization techniques, mitigation methods, and computational approaches introduced in this study reposition device noise from a limitation into a strategic design asset. These principles offer a practical route toward developing next-generation low-power, high-density, and high-performance electronic systems, and they are broadly applicable beyond ferroelectric memories to other emerging electronic devices exhibiting pronounced LFN.