서지주요정보
Studies on charge trap memristor based bio-inspired neuromorphic devices = 전하 트랩 멤리스터 기반의 생체모방 뉴로모픽 소자 연구
서명 / 저자 Studies on charge trap memristor based bio-inspired neuromorphic devices = 전하 트랩 멤리스터 기반의 생체모방 뉴로모픽 소자 연구 / Geunyoung Kim.
발행사항 [대전 : 한국과학기술원, 2024].
Online Access 원문보기 원문인쇄

소장정보

등록번호

8042583

소장위치/청구기호

학술문화관(도서관)2층 학위논문

DMS 24022

휴대폰 전송

도서상태

이용가능(대출불가)

사유안내

반납예정일

리뷰정보

초록정보

With the increasing volume of data that needs to be processed, research on neuromorphic devices is emerging that mimic energetically efficient biological systems. From a hardware perspective, recent studies aim to replicate the parallel information processing characteristics of the central nervous system capable of efficient processing for large amounts of information. Concurrently, active research seeks to emulate complex biological features using simple electronic components. Although various next-generation semiconductor devices are being explored to realize this, challenges such as leakage currents, device variability, and reliability issues persist for the large-scale integration. Continuous research efforts are necessary to address these issues. In this study, I aimed to improve the limitations of existing memristor-based next-generation semiconductor devices based on the charge trap mechanism. Through an examination of the charge trap-based operational mechanism and band engineering of the oxide thin film layer, I successfully implemented stable characteristics in artificial synapse components. Furthermore, by implementing a biomimetic artificial nociceptor with adjustable thresholds for the first time, I aimed to present a new direction in the research of next-generation neuromorphic devices.

인공지능 학습과 같이 처리해야하는 데이터 양이 증가함에 따라 에너지적으로 고효율 적인 생물학적 시스템을 모사하는 뉴로모픽 소자 연구가 대두되고 있다. 하드웨어 관점에서의 연구에서는 중추신경계의 병렬적인 정보처리 특성을 모사하여 대용량 정보를 에너지 효율적으로 처리할 수 있는 컴퓨팅 플랫폼을 구현하고자 하는 연구와 함께, 실제 복잡한 생물학적인 특성을 간단한 전자소자를 통해 모사하고자 하는 연구 또한 활발히 이루어지고 있다. 이를 구현하기 위한 다양한 차세대 반도체 소자들이 연구되고 있으나, 대규모 집적에 따른 누설전류 문제와 소자간 산포도 문제, 신뢰성 문제 등이 남아있어 이를 해결하기 위한 지속적인 연구가 필요하다. 본 연구에서는 전하 트랩 메커니즘 기반의 멤리스터 소자를 기반으로 기존 멤리스터 기반 차세대 반도체 소자의 한계점들을 개선하고자 하였다. 전하 트랩 기반 동작 메커니즘에 대한 고찰과 함께 산화물 박막층의 밴드 엔지니어링을 통하여 안정적인 인공 시냅스 소자 특성을 구현하였고, 역치 조절이 가능한 생물학적인 통각수용기 모사 소자를 최초로 구현하여 차세대 뉴로모픽 소자 연구의 새로운 방향성을 제시하고자 하였다.

서지기타정보

서지기타정보
청구기호 {DMS 24022
형태사항 ix, 94 p. : 삽도 ; 30 cm
언어 영어
일반주기 저자명의 한글표기 : 김근영
지도교수의 영문표기 : Kyung Min Kim
지도교수의 한글표기 : 김경민
수록잡지명 : "Threshold Modulative Artificial GABAergic Nociceptor". Advanced Materials, v.35, no.47, (2023)
수록잡지명 : "Retention Secured Nonlinear and Self‐Rectifying Analog Charge Trap Memristor for Energy‐Efficient Neuromorphic Hardware". Advanced Science, v.10, no.3, (2022)
Including appendix
학위논문 학위논문(박사) -- 한국과학기술원 : 신소재공학과,
서지주기 References : p. 85-91
주제 Memristor
Charge trapping mechanism
Artificial synapse
Artificial nociceptor
Neuromorphic computing
멤리스터
전하 트랩 메커니즘
인공 시냅스
인공 통각수용기
뉴로모픽 컴퓨팅
QR CODE

책소개

전체보기

목차

전체보기

이 주제의 인기대출도서

Neuromorphic computing architecture with artificial synapses.[11] a) Conventional von Neumann architecture with 'memory bottleneck*issue. b)Program-storage computing structure. c)Neuro-inspired hardware architecture. d) Neuro-inspired parallel computing with crossbar array structure. e) Biological brain schematic. f) Biological neurons and synapses schematic forsynaptic signals integration and fir

Active array structure example with Ta/HfO./Pd memristor and transistor integration (1TIR) process. [21] For the ITIR array structure, rows share bottom electrode lines while columns share top electrode lines and gate lines for gate contact.

Passive array structure example with Al20y/TiO2x switching layer based memristor process. [27] a-h) fabrication process for 64x64 passive array device. i-1) Scanning electron microscopy (SEM) images of the patterned array structure without transistor integration.

Leakage current issue schematic for the integrated passive array structure.

Memristor based artificial neuron devices.[p끼]I-V characteristics ofa)NbO2, b)B-Te,and c)Ag/Hf0 based TS devices.

Illustration of a biological neuron system with series resistor (Ra), parallel capacitor (Cm) and TS [35] memristor device.

Illustration ofa CTM device with charge trap-based operation mechanism.

Research trends in the charge trap mechanism based memristor.(+2-40)

The cross-section transmission electron microscopy (TEM) image of the device and fast-Fourier transform (FFT)image.

The x-ray diffraction results ofTazOs, NbzOs-x and Al2O3-y films on a SiO2 substrate. All thin films were amorphous as deposited.

Analog I-V characteristics ofthe PTNAT device. a) The resistance switching I-V curves ofthe device with 1 nA compliance current (Icc) measured at different positive voltage sweeps (6 Vto 10 V) and -10 V fixed voltage sweeps. b)I-V curves with continuously increasing voltage sweeps (6 V to 10 V) without reset.

Memory margin and uniformity ofthe PTNAT device. a) The on/offratio and the rectifying ratio as a function ofthe applied voltages. b) The cumulative probabilities of the current levels ofLRS in the positive voltage region, which were obtained from the first voltage sweeps in 20 cells.

The endurance properties up to 105 cycles read at4 V at room temperature.

Synaptic characteristics ofthe PTNAT device. a) The average and standard deviation ofboth LTP and LTD for 70 pulses during ten cycles. b) 10cycles ofLTP and LTD operation ofthe device.

The retention characteristics ofthe PTNAT device of3-bit intermediate states at 125 oC at2V.

Pt/NbzOsu/ii(PNT) singlelayer deviceI-V curves (resetvoltage of-lVto-4V). The device showed memory operation only with the single Nb2Os-xlayer.

The resistance switching I-V curves ofthe a) Punbz0ss/Alb0./Ti(PNAT), b) Pt/Taxoy/bbyOss/T (PTNT),and c) Pt/TagOs/NbzC5.x/Alzos.Ti (PTNAT) device. EachI-V curve was measured with a 10V positive voltage sweep and a -10V negative voltage sweep with a 10-6 A compliance current.

The retention characteristics ofa) PNAT (4Vread),b)PTNT (3 V read), and c)PTNAT (4V read) at room temperature condition.

The stable retention characteristic ofPTNAT at the harsh temperature condition (2 V at 150 'C).

REELS and UPS bindingenergy cutoffresults ofa)AlzO3y, b)NbzOs-x, and c) TazOs. Used equation for the X value extraction is as follows:

Schematic energy band diagram models ofa) PNAT, b) PTNT, and c) PTNAT devices.

a) HRS fitting (P-F emission) with -5 V reset state ofthe PNT device. b) Temperature dependency ofthe HRS from 50 oC to 90 'C condition. Each curve includes 3 cycles of measurements (light colors), and the average value wasplotted with emphasized lines. c) Arrhenius fittingresult from the temperature dependent HRS The fitted voltage region was 0.8 Vto 0.9V, and the fitted activation energy was aro

Pu/Tatop/AloosaTIt (PTAT) double layer device I-V curves. The device showed no stable memory operation.

a)I-V curves ofthe LRS measured attemperatures ranging from 40 to 70 iC. b)Schottky emission form (ln(J/T2) VS. E1/2) plot for the voltages ranging from 3.5 to 6.5V in LRS with the calculated TazOs layer partial electric field. c) Arrhenius fitting result from the temperature dependent LRS. The fitted voltage region was 3.5 Vto 6.5 V, and the fitted activation energy was around 0.6 eV that indicat

a) Analog LRS states Schottky emission fitting in +4.5 V read voltage and room temperature condition after+6 Vto +9V different sweep voltages. b)High-frequency permittivity (8op) and c) Schottky barrier height (0B) extracted at each analog state. 1BB decreased with a more programmed analog state (higher conductance state).

Conduction mechanism fittingresults limited by different oxide layers. Fittingresults ofa) Schottky emission with TazOs partial field, b) Poole-Frenkel (P-F) emission with NbzOs-x partial field and c) Schottky emission with AlzO3-ypartialfield calculation. Dueto the highaffinity value ofNbzOs-x layer(2~4), bulklimited mechanism temperature dependent P-F emission) was assumed. The calculatedCop val

XPS depth profilingresults ofa)Ta4f,b)NbComorc)Al2pcorelevels ofthe TazOs(10nm)/NbzO x(28nm)AlzO3,,(8 nm)/Tisample. The leftpanels show the raw XPS depth profile data. The right panels enlarge the square region.

Nb 3d XPS spectra deconvolution results of PTNAT device. Nb 3d XPS depth profile from a TazOs/NbzOsa interface (3~19 etch level) to b)NbzOss/AllOoy interface region (27~43 etch level). In the TazOs /NbzOs-x interface region, Nb2Ospeak ratio increased due to the oxidization during the PEALD process. As the etch level increased near to Nbz0ss/Alz0s.y interface region, various NbzOs-x sub-phases were

Al2pandTi2pXPS spectra deconvolution results ofPTNATdevice. Al2p XPS depthprofilefrom a) NbzOss/Al20s.g interface (27~43 etch level) to b) Al203-y/Ti interface region (45~59 etch level). The Al 2p peak shifted toward lower binding energy at the interface region. c) Ti 2p XPS depth profile result shows additional TiO2 film generated at the Al2O3-y/Ti interface.

Reverse structure (Ti top electrode & Pt bottom electrode) CTM device operation. a) Ti/A12O3. y/NbzOsiatPtdevice andb) Ti/AlzOsy/Nbz05.x/2a205.2tdevice showed bi-directional set operation. c) Ti/NbzOs /Altosw/acoy/Ptdevice showed rectifying behavior.

The schematic energy band diagram ofthe device for zero bias condition. Shallow and deep trap levels exist in the non-stoichiometric Nb2Os-x layer.

Theillustration oftheresistance switchingprocess with chargetrappingand de-trapping. Inthe HRS, deep traps are empty, and conduction is not fluent with the high band offset at the TazOs interface (I). When a positive setbias is applied on the Ptelectrode, the electrons can be trapped inthe NbzOs-xdeeptrap sites (II). With the induced negative space charge, the partial electric field across the Taz

A top-view SEMimage ofthe 32x32 PTNAT MCA.

a) I-V curves ofrandomly chosen 80 cells with the halfvoltage scheme. The inset shows the half voltage array measurement scheme. b) Thecumulative probabilities ofthe current levels at the LRS(read voltages of1.3V,4V,and 6 V), and the HRS (read voltages of4V and6 V).

The serial (left) and parallel (right) programming schemes for weight update.

Suggested potentiation and depression scheme for parallel programming.

The double-layer perceptron neural network used for the MNIST dataset training simulation. It consists of784 input neurons, 256 hidden neurons, and 10 output neurons.

Memristive neural network simulation for MNIST classification. a) Schematic diagram ofdouble- layer perceptron neural network. b) Flow chart ofthe training process. After the backpropagation, output value (y), error value (6),and desired weightupdate value (AW) were calculated. Asingleprogramming pulse based on the synaptic property was applied to the selected cell for the weightupdate.

MNIST recognition test fitting parameter ofPTNAT device. a) Synaptic plasticity characteristic 0 PTNAT device. b)Fittingparameter from the synaptic plasticity result. Used equation for the fittingis as follows

a) The MNIST data recognition rate by training epoch considering the experimental synaptic characteristics ofthe CTM device. The final accuracy was about 91%. b) Estimated energy required for training by serial and parallel programming in halfvoltage scheme and those in the suggested scheme.

Operation scheme comparison of3-terminal flash structure and 2-terminal CTM device.「2]

a) I-V curve comparison of the PTNAT device with 3 nm and 5 nm AlzO3-y layer thickness. b) Retention tendency of the PTNAT device with 3 nm, 5 nm and 8 nm device. With thinner tunneling layer, operation voltage decreased, but the memory stability degraded.

a) Schematic image ofthe 3-layer CTM device. b) Schematic image ofthe NP-layer inserted CTM device.

a) The cross-section TEM image ofthe Au NP-CTM device. The thickness ofthe Au nanoparticle layer was around 5 nm. b) A top-view SEM image ofthe Au NP layer. The radius ofthe Au nanoparticles was around 5 nm.

AFM 3D topographyimage ofthe Au NP-CTM device.

Top-view SEM images ofthe Au NPlayer with a)200 seconds deposition (200 'Cannealing), b)200 seconds deposition (no annealing), c) 100 seconds deposition (200 0C annealing), and d) 100 seconds deposition (no annealing) conditions.

a)I-V characteristic comparison ofthe reference PTNAT device (black) and the Au NP-CTM device (red). b) The analog switching I-V curves ofthe device with various Icc conditions (5 nA to 10DA) measured at fixed 8 V positive voltage sweeps after -8 V reset.

The on/off ratio and the rectifying ratio as a function ofthe applied voltages ofthe Au NP-CTM device.

Analog reset characteristic comparison for the different Au NP layerinsertion position, which are a Al2O3-y tunneling layer interface, b) Nb2Os-x charge trap layer bulk, and c) Ta2Os blocking layer interface.

Single programmingpulse response ofthe reference PTNAT device.

Single programming pulse response ofthe Au NP-CTM which the Au NP layeris deposited at a) AlzO3-ytunneling layer interface, b) Nb2Os-x charge trap layer bulk, and c) TazOsblockinglayer interface.

Potentiation characteristics ofthe reference PTNAT CTM device fora) 11 Vamplitude and b) 12V amplitude. Thepulse width conditions were 10 us~ 500 us persinglepulse input. The read pulse amplitude was 4V.

Potentiation characteristics ofthe Au NP-CTM device fora) different amplitude conditions (7V~ 10 V) for 10 us width and b) 10V amplitude condition with different width conditions (100 ns~20ps). Theread pulse amplitude was 3 V.

10 cycles ofLTP and LTD operation ofthe Au NP-CTM device.

Au NP-CTM memory stability characteristics. a) The endurance properties up to 105 cycles read at 3 V for room temperature. b) The retention properties up to 104 seconds at2.5 V for 125 으C condition.

COMSOL based electrical field simulation results for a) vertical and b) lateral reference line

a) I-V curves of the Au NP-CTM LRS measured at temperatures ranging from 40 to 70 b) Arrhenius fitting(ln(J/T2) VS. q/kT) result from the temperature dependent LRS. The fitted voltage region was 2.6 V to 3.6 V.c) Fitted activation energy results which were around 0.59 eV, that indicates the Schottky barrier height at the TazOs / NbzOs-x interface analyzed in Chapter 2.

Potentiation comparison for 8 V amplitude condition in Au NP-CTM device and 13 V amplitude condition in PTNAT device.

The illustration ofthe charge trapping process comparison ofthe PTNAT device and AuNP-CTM device. Reference device has slow trapping process at shallow trap to deep trap level. However, with the Au NP trap site (~5.1 eV), direct trappingat Nb2Os-xdeeptrap level can be occurred which enhance trapping probability. Also, partial field enhancement accelerates charge trapping process ofthe Au NP-CTM de

Illustration ofthe Au NP-CTM device for higher programming efficiency.

GABAergicnociceptor configuration. Whennociceptors generatepain signals, theyrelease glutamate neurotransmitters onto the postsynaptic neuron. Thebindingofglutamate to AMPAand NMDAreceptors onthese neurons causes excitatory postsynaptic potential (Depolarization) and propagates the pain signal to the nervous system when it exceeds the threshold (Vn). Meanwhile, the inhibitory interneuron releases

Artificial GABAergic nociceptor structure and modulative threshold characteristics depending on priming voltage conditions. a) TEM image ofthe artificial GABAergic nociceptor device and the FFTimage. b 30cyclesofI-V curves comprising three steps; priming (0V→2V→0V, 1),switching (0V→11V→0V 2 and 3),and reading (0V→-11V→0V, 4 and 3).c) 100 cycles ofI-V curves under 0 V (black curves) 2 V(red curves)

The I-V characteristics ofthe Pt/Ti/NbzOsaxAlzOs.y/Pt/Ti device with a single oxidant Al2O3-ylayer. a) Al2O3-ylayer with an O3 oxidant (Pt/Ti/24 nm Nb2Os.x/7 nm o-AL20s,p/Ti) and b) Al2O3-ylayer with a H2O oxidant (Pt/Ti/24 nm Nb2Os.x/7 nm h-Alz0ps/Pt/Ti).

The I-V switching characteristics with various o-AlzO3-y and h-AlzO3-y thickness combinations for identifying the roles ofeach layer. a, b)Athinnerh-Al2O3-y cases, 4 nm o-AlzO3.y/4 nmh-AlzOs.y (a) and 6 nm O-

Device-to-device variability. The cumulative probabilities ofthe Vth values at each VPRM condition obtained from 20 devices. The dashed lines are the average Vth values.

XPS depth profile results and operation model for the artificial GABAergic nociceptor. Dept profiling results ofa)Nb 3dcore levels atthe Nb2Os-xlayer andb) 0 1s core levels atthe Al2O3-y doublelayer. C The energy band diagram ofthe GABAergic nociceptor device for zero bias condition. Two types oftrap level are shown, which are the Vo traps in NbzOs-xlayer (trap sites 1) and the H: traps inh-AlzO3.

XPS spectra deconvolution results and peak area ratio results. a, b)XPS spectra deconvolutionre ofNb3d(a)and 0 1s(b) peaks. C.d.d) The deconvoluted peak area ratio in the NbzOs-xlayer(c) and inthe A y layer (d). The surface of Nb2Os-x is exposed to the air and oxidizes, resulting in the detection ofmore 1 However, in actual devices, it can be ignored because the top electrodes cover the surface of

Conduction mechanism P-F conduction fittingresults. a) In(1/(ExT52)) VS. 1/T plot for temperature- dependentP-F fittingofthe low resistance state programmed byIcc = 10-6A, with temperature ranges from 60 'C to 100 'C. b) ln(J/E) VS. E1/2 plot for the different Icc. The voltage range is from -4 V to -5 V. The electric field on the Nb2Os-xlayer was calculated considering the thickness and the dielec

Modulative threshold triggering demonstration for various VPRM conditions of0V (black),2V (red). 4V (blue), and 6V (green).

Modulative relaxation demonstration with various duration of2 V GABA pulses, 100 ns (red), 1 fts (blue), and 10 us (green). No GABA pulse results are also included for reference (black).

a)Current responses for no injury condition (black), after injury condition (red),and modulated after injury conditions with3 V(blue),4V (green), and5V (purple)curingpulsesignals. Each curingpulsesignal was appliedbetween the serial pulse trains. b) The average output currents were plotted at each inputpulse amplitude, which shows the modulative allodynia and hyperalgesia behaviors ofthe device.

No adaptation behavior demonstration. Even when the noxious inputs are introduced continuously, a nociceptor system should respond consistently to avoid damage from the repeated noxious stimuli. The device generated continuous outputs under the continuous stimulation for 100 input pulses (-6 V, 140 ms), confirming that the device exhibits no adaptation behavior.

Concept illustration ofthe hot- and cold-sensitive thermoreceptors in the human skin

Temperature-dependent I-V curves ofb) 4V and c)0VVPRM condition.

a) Hot-sensitive thermoreceptor demonstration with the 4 VVPRMcondition device applying a 5. V input pulse. The device generated output currents only at a noxiously high temperature condition (70 'C). b Cold-sensitive thermoreceptor demonstration with the 0 V VPRM condition device applying4.0Vinputpulse. The device only reacted with a low temperature condition (20 'C). Utilizing both demonstration

Temperature-dependent I-V curves ofthinner h Al2O3-ylayer (4 nm) device for b)4V and c)0V VPRM condition.

Reaction temperature range changed with thinner h Al2O3-ylayer (4 nm) device. a) Hot-sensitiv thermoreceptor demonstration with the 4 V VPRM condition device applying a 5.5 V input pulse. The devic generated output currents at both warm condition (40 'C) and noxiously high temperature condition (70 'C). 1 Cold-sensitive thermoreceptor demonstration with the 0 V VPRM condition device applying4.5 Vi

Arrhenius plot results ofhigh GABA and low GABA states. The Arrhenius plot (In(AVin) VS. 1/T) for a)4 V VPRM and b)0 V VPRM conditions. The extracted Ea values were 0.59 eV for4 V VPRM condition (hot- sensitive receptor) and 0.35 eV for 0 V VPRM condition (cold-sensitive receptor). These Ea values refer to the energy barrier for de-trapping process (4 V VPRM) and for the trapping process (0 V VPRM

A band diagram model for the thermoreceptor behaviors. The band diagrams in various temperature

Benchmarking of the thermoreceptor performances. This benchmarking table includes sensing performance factors, whose descriptions are as follows. i) Output source: The type ofsensing output, whether it is voltage, current, or frequency. ii) Range: The operating temperature range. iii) Sensing ratio: The ratio ofthe sensing outputto the no response output. iv) Response time: The time delay for the