An artificial nerve system emulating the tactile perception capabilities of human skin holds immense potential for advanced applications in a hyperconnected society, including prosthetic devices, virtual reality interfaces, and intelligent sensors. This system integrates a pressure sensor, analog-operable memory, and a signal processing unit, mirroring the sophisticated functionality of biological neural networks. Despite decades of remarkable advancements in semiconductor-based memory and transistor technologies, sensors for detecting external stimuli predominantly rely on micro-electromechanical systems, which present significant challenges in achieving miniaturization and high-density integration. Our lab introduces an innovative monolithic three-dimensional integrated artificial nerve system, seamlessly combining a self-powered piezoelectric pressure sensor with neuromorphic memory fabricated using ferroelectric hafnium zirconium oxide (HfxZr1-xO2) films. This cutting-edge approach aims to redefine the development of tactile perception technologies, paving the way for transformative breakthroughs in diverse fields.

The number of nodes commonly used in sensory networks is rapidly increasing, which increases the bottlenecks that occur during data communication. To efficiently process large amounts of data and reduce power consumption for data processing, computational approaches that perform preprocessing in sensory networks must be developed. This approach can reduce redundant data movement between sensing and processing devices. In our lab, we consider proximity sensors and intra-sensor computing, where the computational tasks are partially shifted to the sensory terminal.

The implementation of artificial optic nerves is a critical research area for mimicking biological neural networks and advancing neuromorphic computing and biomedical applications. In this study, we propose a novel artificial optic nerve architecture that integrates photodiode synapses with three-terminal Anti-ferroelectric-FET neurons, enabling the continuous optical signal to be dynamically converted into spike signals proportional to light intensity. This system leverages photodiodes for analog sensing and signal processing, offering a natural sensory response, while ferroelectric-MOSFET neurons efficiently generate spike signals. Notably, the proposed architecture supports three-dimensional (3D) integration, overcoming limitations in area and density constraints, thus enabling high-density and compact designs. Our developed artificial optic nerve demonstrates enhanced spatial and energy efficiency, providing a promising solution for neuromorphic computing and bio-interface applications.

We focus on in-sensor computing, integrating various sensors, such as piezoelectric and optical sensors, with semiconductor devices to enable low-power, low-level data processing within sensor terminals. Specifically, we have developed a piezoelectric sensor using AlN material that can detect pressure in real time while achieving a high output voltage exceeding 3 V without requiring additional amplifiers. In addition, we are working on integrating these AlN piezoelectric sensors into arrays with semiconductor devices like FeFETs. Traditional sensor systems often face limitations such as high power consumption and processing delays due to reliance on external data transmission and processing. By leveraging the unique properties of ferroelectric semiconductor devices, we aim to overcome these challenges by enabling real-time data preprocessing and simple computations directly within sensor terminals. This approach not only enhances power efficiency and data transmission speed but also facilitates real-time data processing for effective high-level data processing in applications such as flexible tactile systems, IoT, and autonomous driving.

Combining memory and ROIC (Read-Out Integrated Circuit) with voltage obtained through a piezoelectric sensor enables sensor computing, where data for a neural network can be processed directly within the device. Specifically, an AlN (Aluminum Nitride) sensor, which offers a high piezoelectric voltage due to its unlimited thickness scalability, serves as a core component. This high piezoelectric voltage writes the resistance state of a memory device with continuously variable resistance, such as an FTJ (Ferroelectric Tunnel Junction). The voltage is then distributed according to the resistance state and subsequently transmitted to a Voltage-Controlled Ring Oscillator (VCRO). The VCRO's output, representing frequency changes in response to input voltage, can be used as an input for a Spiking Neural Network (SNN), where it encodes spike signal frequency variations. If made flexible, such a system could be applied to applications like electronic skins (E-skins) and robotics, enhancing real-time processing and reducing data latency in neural networks. This innovation allows for faster and more efficient sensor-based computations in real-world scenarios.