서브메뉴
검색
Micro/Nano Fabrication of Layered Semiconductor Devices for Hardware Implementation of Neuromorphic Computing
Micro/Nano Fabrication of Layered Semiconductor Devices for Hardware Implementation of Neuromorphic Computing
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103644
- ISBN
- 9798314874394
- DDC
- 620.11
- 저자명
- Chen, Mingze.
- 서명/저자
- Micro/Nano Fabrication of Layered Semiconductor Devices for Hardware Implementation of Neuromorphic Computing
- 발행사항
- [Sl] : University of Michigan, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 151 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-11, Section: B.
- 주기사항
- Advisor: Liang, Xiaogan.
- 학위논문주기
- Thesis (Ph.D.)--University of Michigan, 2025.
- 초록/해제
- 요약Artificial intelligence (AI) has been extensively used in the routines of human life and shown superior efficiency in various tasks, including pattern classification, voice recognition, and language processing. However, to solve the complicated problems involving large amount of spatiotemporal data (e.g., controlling of dynamic systems), AI development is approaching the fundamental limitations of the current computing systems based on von Neumann's architecture. To overcome such technological barriers, new neuromorphic computing techniques have been being actively explored, seeking to substantially improve the energy efficiency for computing tasks and extend AI capacities. In addition, such neuromorphic computing techniques mimic the architecture and functioning of the biological neural networks and they can enable in-memory computing architectures with parallel-processing capability.To specifically improve the energy- and time-efficiency of neuromorphic computing, a range of physical systems have been proposed and investigated to serve as the computing units dedicated to hardware-based neuromorphic systems. Among these physical systems, memristors, a new class of electronic devices, have attracted tremendous attention because of their great potential for enabling low-power in-memory computing processes. However, to ultimately realize the memristor-based neuromorphic computing devices and systems for practical applications, we need to address a series of important challenges: (i) Salable nanofabrication methods for producing high-quality memritive materials are still deficient; (ii) The device physics about the unique electronic properties of memristive materials is not fully understood; and (iii) system-level integration of memristor-based neuromorphic computing devices is not fully developed.The presented dissertation projects aim to address part of the aforementioned challenges and realize the following objectives: (1) Development of a scalable nanofabrication approach, termed rubbing-induced site-selective deposition (RISS), capable of producing patterned 2D-material-based memristive device structures without additional lithographic processes; (2) Systematical investigation of the switching characteristics and mechanisms of RISS-produced Bi2Se3 memristors for neuromorphic computing applications; (3) Development of Bi2Se3-based memristive devices capable of directly extracting spatiotemporal information from analogue video signals that could be utilized for computer vision applications ; (4) Creation of Bi2Se3 memristive networks that can realize hardware implementation of neuromorphic computing frameworks for robotic vehicle control.The first part of the thesis presents an advanced nanofabrication technique, termed rubbinginduced site-selective deposition (RISS), which is capable of producing patterned 2D-layered material arrays (e.g., MoS2 and Bi2Se3) without resist-based lithography and plasma etching processes. This method can generate arrays of microscale 2D device features with a high fidelity to the designed patterns and high yield (95%) throughout centimeter-scale areas. Several critical components and factors that notably affect the final yield of RISS processes have been systematically investigated and optimized to enable scale-up device applications. The second part of the thesis presents experimental and theoretical simulation works on the unique bipolar resistive switching characteristics of RISS-produced Bi2Se3 memristors. Specifically, such a Bi2Se3 memristor exhibits a reliable dependence of conductance modulation on the duty cycle of programming pulses, a fast relaxation behavior and a low threshold field for activating memristive switching (104 to 105 V/cm). Such an observed switching characteristic (i.e., short-term memory behavior) is attributed to the field-driven drift and concentration-controlled diffusion of selenium vacancies in Bi2Se3 quintuple layers, as implied by the theoretical simulation result. Based on the aforementioned memristive switching characteristics, we utilize the output of such a device in response to analogue video scanline signals to extract simple graphic information components. We have further developed a hardware-based graphic motion sensor which can generate control command signals in response to analogue video signals with time resolution of 64μs. The fourth part of the thesis presents the construction and implementation of a neuromorphic computing system capable of processing analogue sensory signals with temporal information to generate output signals for controlling dynamic systems. Such a system features a RISS-produced memristive network that can serve as a hardware representation of the reservoir computing framework. The implementation of such a system in robotic vehicle control processes result in three orders of magnitude lower power consumption (12.5µW) in comparison with the software counterpart. In two experimental tasks, this hardware-based neuromorphic computing system exhibits a highly consistent performance in comparison with the traditional digital controllers from which this neuromorphic system learns the control functionality. The low prediction error of the hardware-based neuromorphic computing system is quantitatively indicated by relatively low normalized-root-mean-square-error (NRMSE) values of 0.11 and 1.25 for rover navigation and lever balancing tasks, respectively. The presented device structure and system framework could be further leveraged for constructing low-power hardware-based edge computing and controlling systems for a broad range of miniature robotic systems.
- 일반주제명
- Materials science
- 일반주제명
- Mechanical engineering
- 일반주제명
- Computer engineering
- 키워드
- Nanofabrication
- 기타저자
- University of Michigan Mechanical Engineering
- 기본자료저록
- Dissertations Abstracts International. 86-11B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017358100
■00520260202103644
■006m o d
■007cr#unu||||||||
■020 ▼a9798314874394
■035 ▼a(MiAaPQ)AAI32092586
■035 ▼a(MiAaPQ)umichrackham006120
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a620.11
■1001 ▼aChen, Mingze.
■24510▼aMicro/Nano Fabrication of Layered Semiconductor Devices for Hardware Implementation of Neuromorphic Computing
■260 ▼a[Sl]▼bUniversity of Michigan▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a151 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-11, Section: B.
■500 ▼aAdvisor: Liang, Xiaogan.
■5021 ▼aThesis (Ph.D.)--University of Michigan, 2025.
■520 ▼aArtificial intelligence (AI) has been extensively used in the routines of human life and shown superior efficiency in various tasks, including pattern classification, voice recognition, and language processing. However, to solve the complicated problems involving large amount of spatiotemporal data (e.g., controlling of dynamic systems), AI development is approaching the fundamental limitations of the current computing systems based on von Neumann's architecture. To overcome such technological barriers, new neuromorphic computing techniques have been being actively explored, seeking to substantially improve the energy efficiency for computing tasks and extend AI capacities. In addition, such neuromorphic computing techniques mimic the architecture and functioning of the biological neural networks and they can enable in-memory computing architectures with parallel-processing capability.To specifically improve the energy- and time-efficiency of neuromorphic computing, a range of physical systems have been proposed and investigated to serve as the computing units dedicated to hardware-based neuromorphic systems. Among these physical systems, memristors, a new class of electronic devices, have attracted tremendous attention because of their great potential for enabling low-power in-memory computing processes. However, to ultimately realize the memristor-based neuromorphic computing devices and systems for practical applications, we need to address a series of important challenges: (i) Salable nanofabrication methods for producing high-quality memritive materials are still deficient; (ii) The device physics about the unique electronic properties of memristive materials is not fully understood; and (iii) system-level integration of memristor-based neuromorphic computing devices is not fully developed.The presented dissertation projects aim to address part of the aforementioned challenges and realize the following objectives: (1) Development of a scalable nanofabrication approach, termed rubbing-induced site-selective deposition (RISS), capable of producing patterned 2D-material-based memristive device structures without additional lithographic processes; (2) Systematical investigation of the switching characteristics and mechanisms of RISS-produced Bi2Se3 memristors for neuromorphic computing applications; (3) Development of Bi2Se3-based memristive devices capable of directly extracting spatiotemporal information from analogue video signals that could be utilized for computer vision applications ; (4) Creation of Bi2Se3 memristive networks that can realize hardware implementation of neuromorphic computing frameworks for robotic vehicle control.The first part of the thesis presents an advanced nanofabrication technique, termed rubbinginduced site-selective deposition (RISS), which is capable of producing patterned 2D-layered material arrays (e.g., MoS2 and Bi2Se3) without resist-based lithography and plasma etching processes. This method can generate arrays of microscale 2D device features with a high fidelity to the designed patterns and high yield (95%) throughout centimeter-scale areas. Several critical components and factors that notably affect the final yield of RISS processes have been systematically investigated and optimized to enable scale-up device applications. The second part of the thesis presents experimental and theoretical simulation works on the unique bipolar resistive switching characteristics of RISS-produced Bi2Se3 memristors. Specifically, such a Bi2Se3 memristor exhibits a reliable dependence of conductance modulation on the duty cycle of programming pulses, a fast relaxation behavior and a low threshold field for activating memristive switching (104 to 105 V/cm). Such an observed switching characteristic (i.e., short-term memory behavior) is attributed to the field-driven drift and concentration-controlled diffusion of selenium vacancies in Bi2Se3 quintuple layers, as implied by the theoretical simulation result. Based on the aforementioned memristive switching characteristics, we utilize the output of such a device in response to analogue video scanline signals to extract simple graphic information components. We have further developed a hardware-based graphic motion sensor which can generate control command signals in response to analogue video signals with time resolution of 64μs. The fourth part of the thesis presents the construction and implementation of a neuromorphic computing system capable of processing analogue sensory signals with temporal information to generate output signals for controlling dynamic systems. Such a system features a RISS-produced memristive network that can serve as a hardware representation of the reservoir computing framework. The implementation of such a system in robotic vehicle control processes result in three orders of magnitude lower power consumption (12.5µW) in comparison with the software counterpart. In two experimental tasks, this hardware-based neuromorphic computing system exhibits a highly consistent performance in comparison with the traditional digital controllers from which this neuromorphic system learns the control functionality. The low prediction error of the hardware-based neuromorphic computing system is quantitatively indicated by relatively low normalized-root-mean-square-error (NRMSE) values of 0.11 and 1.25 for rover navigation and lever balancing tasks, respectively. The presented device structure and system framework could be further leveraged for constructing low-power hardware-based edge computing and controlling systems for a broad range of miniature robotic systems.
■590 ▼aSchool code: 0127.
■650 4▼aMaterials science
■650 4▼aMechanical engineering
■650 4▼aComputer engineering
■653 ▼aAdvanced nanofabrication
■653 ▼aNeuromorphic computing
■653 ▼aReservoir computing
■653 ▼aArtificial intelligence
■653 ▼aNanofabrication
■690 ▼a0548
■690 ▼a0794
■690 ▼a0464
■71020▼aUniversity of Michigan▼bMechanical Engineering.
■7730 ▼tDissertations Abstracts International▼g86-11B.
■790 ▼a0127
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358100▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


