Ferroelectric-based transistor for neuromorphic system

Abstract

본 논문은 현대 컴퓨팅 시스템이 직면한 폰 노이만 구조의 병목현상을 해결하고자 강유전체 재료를 사용한 신경 모방 시스템의 설계 및 구현에 중점을 둡니다. 데이터 중심 애플리케이션은 빠른 접근 시간, 스케일링 호환성, 저전력, 저비용, AI 작업 실행 능력을 요구하며, 빅데이터의 급격한 증가는 알고리즘 및 아키텍처의 복잡성을 증대시켜 에너지 소비를 크게 증가시키고 있습니다. 이러한 기술적 요구사항을 충족시키기 위해 뇌의 병렬 처리를 모방하는 신경 모방 컴퓨팅 아키텍처에 대한 연구로, CMOS 기술을 사용한 기존의 신경 모방 시스템이 등장하였습니다. 하지만, CMOS 기술 기반의 시냅스 회로는 매우 복잡하여 넓은 면적과 제작 비용 상승을 요구하는 단점이 있습니다. 이러한 단점으로, CMOS 회로 기반 시냅스는 실제 인간 뇌를 모방하기에 부적절 합니다. 이러한 한계를 단일 시냅스 소자를 통해 해결하고자 합니다.

단일 시냅스 소자의 대표적인 강유전체 재료는 전력 없이도 데이터를 유지할 수 있는 능력으로 인해 신경 모방 애플리케이션에 매우 유망합니다. 이 재료들은 반복적인 데이터 읽기 및 쓰기를 지원하며, 다중 수준 저장과 빠른 스위칭을 가능하게 하여 복잡한 계산과 실시간 처리에 적합합니다. 그러나 이 재료들은 극성 특성을 위한 고온 처리 요구로 인해 현재 제조 공정 및 CMOS 기술과의 호환성 문제에 직면하고 있습니다. 본 논문은 다양한 재료와 구조를 사용하여 강유전체 메모리를 활용한 시스템을 구축하는 것에 중점을 두며, 이를 통해 뇌에서 영감을 받은 뉴로모픽 컴퓨팅 기능을 구현합니다. 다양한 강유전체인PVDF, PZT, HZO 등 재료의 특성을 비교 분석하며, 각 재료가 신경 모방 장치에 적합한지에 대한 평가를 진행합니다. 최적화된 강유전체 공정을 가지고 CMOS 기술과의 호환성 및 스케일링 가능성을 평가하고, 이를 통해 신경 모방 아키텍처의 효율성을 검증합니다. 추가적으로, 다양한 재료들의 전기-기계적 특성 및 처리 호환성을 상세히 검토하여, 신경 모방 장치에서의 최적화 사용 방안을 제시합니다.

실리콘 채널 기반 FeFET 및 산화물 채널 기반 Fe-TFT의 개선을 통해 신경 모방 시스템의 효율성과 신뢰성을 높이는 방안을 탐구합니다. 퍼니스, 빠른 열처리 및 레이저 어닐링 기술을 사용하여 다양한 물질이 사용되는 복합 공정에서의 재료 특성을 최적화하고, 이를 통해 메모리 장치의 성능을 향상시키는 방안을 모색합니다. 이러한 기술을 통해 논문에서는 현대 컴퓨팅 시스템의 진화에 따른 주요 도전과제를 다루며, 이를 극복하기 위한 강유전체 메모리 기술의 접근법과 다양한 강유전체 기반 트랜지스터 구조를 탐구합니다. 첫번째로, 실리콘 채널 기반 FeFET의 열처리 과정에서 오는 확산 문제를 제어하기 위하여 기능성 확산 차단막을 활용한 메모리를 제시합니다. 더 나아가, 근본적인 확산 문제를 해결 하기 위해 산화물 채널 기반 bottom gate Fe-TFT를 통해 메모리 기능을 검증합니다. 마지막으로, 시스템 통합 과정에서 엄격한 열 에너지 관리를 위해 레이저 어닐링 기술을 적용한 Fe-TFT 소자를 검증합니다. 본 연구는 강유전체 재료를 사용한 신경 모방 시스템의 성능과 신뢰성을 향상시킬 새로운 방법을 제시하며, 향후 지능형 컴퓨팅 솔루션으로의 가능성을 모색합니다. |In the semiconductor industry, data processing performances are growing exponentially under Moore's Law. Despite these advancements, current computing architectures still face increasing technical demands for speed, efficiency, and adaptability. The conventional von Neumann architecture has reached a technical performance limit, because of the serial processing bottleneck between memory and the processor. This paper addresses a technology that emulate the synapse-processing capabilities, which are necessary for exploring neuromorphic computing. Neuromorphic computing promises to overcome significant limitations of current technology by leveraging parallel processing and energy efficiency similar to biological systems, making this exploration critically important. This research focuses on ferroelectric material based transistors employing Lead Zirconate Titanate (PZT) and Hafnium Zirconium Oxide (HZO), which is one of appropriate candidates for developing neuromorphic systems that can effectively emulate cognitive functions such as memory retention and sensory processing without the need for continuous power supply. The materials are particularly advantageous due to their outstanding polarization capabilities, which are essential for multi-level characteristics that excellently mimic the synaptic weights of the human brain, leading to lower power operation. However, typical high temperature annealing processes (above 550°C) may affect the channels and electrodes, limiting their use in various system. Therefore, this study precisely investigated the fabrication, functionalization, and application of these ferroelectric materials within various semiconductor devices, focusing particularly on their integration into ferroelectric field-Effect transistors (FeFETs) and ferroelectric thin-film transistors (FeTFTs). Initially, processes were optimized and the electrical properties of various ferroelectric materials such as PVDF, PZT, and HZO were analyzed. Through the analysis of parameters including dielectric constant, materials suitable for low-power operation and neuromorphic applications were explored. A design engineering with various ferroelectric materials approach was taken to integrate P-Si and HZO ferroelectrics. Instead of adding a simple sub-layer, which is required generally to protect Si diffusion to ferroelectric gate oxide, whereas it causes power consumption and depolarization, we added a functional layer to improve the memory effect compared to the existing one. The proposed device has an in situ HZO/HfO2/Al2O3 stacked film structure, which is compatible for Si with the metal–oxide–semiconductor (MOS) process based on all atomic layer deposition. Since the appropriate bandgap difference between Al2O3 and HfO2, stable charge trap operation was achieved. High-quality ferroelectric HZO film characteristics were shown by minimizing defects and Si diffusion through the sub-layer of Al2O3/HfO2. Therefore, HZO as a blocking layer enhances the memory performance of the charge trap structure due to its specific polarization effect. The proposed device had the high polarization characteristics of HZO (2Pr > 20 μC/cm2) along with a MOS-cap window (> 4 V), good retention capability (> 10 years), fast program/erase response operation times (< 200 μs⁠), and strong durability (> 105 cycles) while operating as a form of single level cell. By comparing Al2O3 and ferroelectric HZO as a blocking layer of the charge trap device, we confirmed that the HZO/HfO2/Al2O3 multi-layer structure had excellent characteristics according to various memory performance indicators. Our proposed high-performance charge trap flash memory can be employed in various applications, including Si-based three-dimensional structures with artificial intelligence systems. Additionally, in FeTFT, neuromorphic devices using PZT and HZO materials were investigated. An indium gallium zinc oxide (IGZO) channel back-gate TFT structure was chosen to address the diffusion of the channel material during the annealing process for crystallization of PZT and HZO. By annealing PZT, post- deposition process using IGZO as channel structure, the co-diffusion of bonding material and oxygen can be minimized, resulting in high and stable performance of FeTFT. During the production of the oxide-based multi- layer structure, the partial pressure conditions of the system were strictly investigated to minimize diffusion and defect phenomena. The basic operations of synaptic short-term memory (STM) and long-term memory (LTM) were also analyzed to confirm the applicability of neuromorphic devices. Due to the high dielectric constant and polarization properties of PZT, the power consumption of the spike signal used for spike-dependent plasticity changes can be reduced to 10 pJ. Additionally, a wide dynamic range of Gmax/Gmin≅ 1000 was obtained, and the channel conductance was maintained for more than 40000 s. The optimized pulse achieved multi-level states (> 32), making the learning process efficient. To solve the more fundamental thermal problem, HZO integrated IGZO TFT as a neuromorphic device through laser process engineering. For selective annealing of hafnia-based FeFETs, a nanosecond pulsed laser was used to precisely control the depth of heat penetration within the thin film. Sufficient thermal energy was transferred to the IGZO oxide channel and HZO ferroelectric gate oxide without causing thermal damage to the bottom layer, which has a low transition temperature (< 250 °C). Using optimized laser conditions, fast response times (< 1 μs) and excellent stability (cycles > 106, retention > 106 s) were achieved in ferroelectric HZO films. The resulting FeFET exhibited a wide memory window (> 1.7 V) along with a high on/off ratio (> 105). It also showed appropriate ferroelectricity (2·Pr, 14.7μC/cm2). Due to its high multi-level state processing and power efficiency, PZT TFT shows recognition accuracy of up to 84.72% in neuromorphic benchmarking systems. On the other hand, HZO FeFET manufactured using laser annealing has a fast response time and excellent stability, making it suitable for real-time processing and various structural configurations, and achieving a recognition rate of over 88.83%. Both structures are promising for neuromorphic systems with high recognition accuracy. These advances highlight the critical role of ferroelectric devices in enhancing cognitive abilities and transforming neuromorphic computing with high accuracy and efficiency. The proposed neuromorphic frameworks improve structural integration and operational efficiency, enhancing density and energy efficiency. Advanced material characterization and device architectures drive next-generation computing technologies. These innovations create efficient, robust systems capable of handling complex computational demands. This work sets a benchmark for future research and the integration of neuromorphic technologies. Keywords: Ferroelectric, Field-Effect Transistors (FET), Low-power system, Neuromorphic computing