187037 research outputs found
Sort by
광전자 촉매 반응을 통한 탄화수소로의 이산화탄소 환원을 위한 기체확산전극의 설계
No full text학위논문(박사) - 한국과학기술원 : 생명화학공학과, 2025.2,[xxx, 390 p. :]In a push toward electrification of the chemical industry to enable closure of the carbon cycle, electrochemical CO2 reduction (ECO2R) is studied with renewed interest over the last two decades. It is identified that increasing the technology readiness level of this method is mainly hindered by its subpar energy efficiency and the stability of the electrodes when attempting to support current densities in excess of 100 mA cm–2. Overcoming these hindrances requires addressing ECO2R on three levels of scale: i) the electrocatalyst, ii) the gas-diffusion electrode, and iii) the electrolyzer system. As a result of this research, it is upheld that under the conditions of interest, it are electrode and system design that dictate the performance metrics of ECO2R, rather than the exact nature of the electrocatalyst. It was found that the stability of ECO2R may be increased by using nonconductive gas-diffusion layers (GDLs), and a new type of GDL based on metal oxide materials is developed herein. In using such GDLs, electrocatalyst morphology remains pivotal for efficient catalyst utilization by ensuring high in-plane conductivity. These three levels of scale are integrated in the final part of this research, in which the concept of photo-assisted ECO2R for enhancing energy efficiency is studied by synthesizing and evaluating plasmonic electrocatalysts.한국과학기술원 :생명화학공학과
NeFL: Nested Model Scaling for Federated Learning With System Heterogeneous Clients
No full textFederated learning (FL) enables distributed training while preserving data privacy, but stragglers-slow or incapable clients can significantly slow down the total training time and degrade performance. To mitigate the impact of stragglers, system heterogeneity, including heterogeneous computing and network bandwidth, has been addressed. While previous studies have addressed system heterogeneity by splitting models into submodels, they offer limited flexibility in model architecture design, without considering potential inconsistencies arising from training multiple submodel architectures. We propose nested federated learning (NeFL), a generalized framework that efficiently divides deep neural networks into submodels using both depthwise and widthwise scaling. To address the inconsistency arising from training multiple submodel architectures, NeFL decouples a subset of parameters from those being trained for each submodel. An averaging method is proposed to handle these decoupled parameters during aggregation. NeFL enables resource-constrained devices to effectively participate in the FL pipeline, facilitating larger datasets for model training. Experiments demonstrate that NeFL achieves performance gain, especially for the worst-case submodel compared to baseline approaches (7.63% improvement on CIFAR-100). Furthermore, NeFL aligns with recent advances in FL, such as leveraging pre-trained models and accounting for statistical heterogeneity.
짝지어진 데이터에 대한 신경 상미분방정식의 시뮬레이션 비의존 학습
No full text학위논문(석사) - 한국과학기술원 : 전산학부, 2025.2,[iv, 27 p. :]In this work, we investigate a method for simulation-free training of Neural Ordinary Differential Equations (NODEs) for learning deterministic mappings between paired data. Despite the analogy of NODEs as continuous-depth residual networks, their application in typical supervised learning tasks has not been popular, mainly due to the large number of function evaluations required by ODE solvers and numerical instability in gradient estimation. To alleviate this problem, we employ the flow matching framework for simulation-free training of NODEs, which directly regresses the parameterized dynamics function to a predefined target velocity field. Contrary to generative tasks, however, we show that applying flow matching directly between paired data can often lead to an ill-defined flow that breaks the coupling of the data pairs (e.g., due to crossing trajectories). We propose a simple extension that applies flow matching in the embedding space of data pairs, where the embeddings are learned jointly with the dynamic function to ensure the validity of the flow which is also easier to learn. We demonstrate the effectiveness of our method on both regression and classification tasks, where our method outperforms existing NODEs with a significantly lower number of function evaluations.한국과학기술원 :전산학부
일반 WSI 및 다중 클래스 신규성 감지를 활용한 비정상 WSI 패치 선택
No full text학위논문(석사) - 한국과학기술원 : 데이터사이언스대학원, 2025.2,[iv, 34 p. :]Whole-slide images (WSIs), as high-resolution scans of tissue samples, are essential for diagnosing and predicting disease outcomes, especially in cancer detection. However, the lack of detailed patch-level annotations poses challenges for WSI classification. Weakly supervised learning models, such as multiple instance learning models (MIL models), address the challenge of relying only on slide-level labels when patch-level labels are not provided. In MIL models, WSI is classified as abnormal if it contains even a single abnormal patch. This characteristic can lead to reduced classification accuracy, especially when abnormal regions occupy a small portion compared to normal regions. This study proposes a method utilizing mixup-based Multi-Class Novelty Detection (Multi-Class ND) to train a Novelty Detection (ND) model using normal WSI patches and remove normal patch instances from the abnormal WSI patch bags. The method removes normal patch instances from abnormal WSI patch bags to improve classification accuracy. Using a colon WSI dataset containing multiple normal and dysplasia classes, the proposed method demonstrated superior classification performance compared to approaches that do not remove normal patches. This methodology can be applied to other WSI datasets containing multiple normal and dysplasia classes.한국과학기술원 :데이터사이언스대학원
하이브리드 위성-지상 중계기 네트워크에서의 보안 빔포밍
No full text학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[ii, 11 p. :]This paper focuses on designing a relay beamformer to maximize a sum secrecy rate in hybrid satelliteterrestrial relay networks. The system employs an amplify-and-forward relay in an uplink multi-user scenario. To resolve the nonconvexity of the formulated optimization problem, slack variables are introduced, and first-order Taylor expansion is applied. The problem is then solved using successive convex approximation. Simulation results compare the sum secrecy rate of the proposed scheme with baselines.한국과학기술원 :전기및전자공학부
고해상도 연속 시간 델타-시그마 변조기에 관한 연구
No full text학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[iii, 33 p. :]This paper presents various techniques to improve high-resolution continuous-time delta-sigma modulator (CT DSM) architectures. First, a novel architecture employing a passive noise-shaping (PNS) successive approximation register (SAR) analog-to-digital converter (ADC) is proposed to address the issue of additional quantization noise introduced by digital noise-coupling (DNC) techniques. Since the loop filter of the PNS-SAR is implemented using charge-sharing, the proposed approach improves resolution with minimal additional power consumption. By leveraging DNC and PNS-SAR, the proposed architecture achieves high resolution with a single operational amplifier (op-amp), minimizing the design complexity. A prototype implemented using a 180 nm CMOS process achieves a signal-to-noise-and-distortion ratio (SNDR) of 100.1 dB over a 50 kHz bandwidth at a sampling frequency of 6.4 MHz. The resulting Schreier Figure of Merit is 178.7 dB, demonstrating competitive performance when compared to other ADCs with similar bandwidths. Next, this work addresses challenges inherent to multi-stage noise-shaping (MASH) structures. MASH DSMs suffer from resolution limitations due to quantization noise leakage caused by mismatches between analog and digital transfer functions. By applying analog noise-coupling (ANC) to shape the quantization noise of the first stage, the proposed approach shifts the effects of quantization noise leakage to higher frequencies, thereby preventing degradation of in-band noise performance. Additionally, a compact architecture is achieved by reusing the ANC buffer for quantization noise extraction. Furthermore, a passive noise-shaping SAR ADC is incorporated in the second stage to enhance resolution while incurring negligible power consumption. This study contributes to the advancement of CT DSM architectures by proposing techniques that improve resolution, robustness, and power efficiency.한국과학기술원 :전기및전자공학부
육방정계 질화붕소 기반 2차원 이종 접합 구조의 원자 결함에 대한 다공간 밀도 범함수론 연구
No full text학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[vii, 51 p. :]In this master's thesis, we analyze atomic defects in two-dimensional heterojunction device structures based on hexagonal boron nitride (hBN) using the newly developed multi-space density functional theory (MS-DFT). First, MS-DFT simulations are performed on vertically stacked tunneling device structures consisting of graphene/hBN/graphene with various atomic defects introduced into the middle hBN layer. Subsequently, the spin component, a quantum mechanical factor, is extensively analyzed when quantum hybridization negative differential resistance (QH-NDR) occurs due to quantum hybridization under source-drain voltage.
Next, Random telegraph noise (RTN) generated by carrier trapping in the hBN substrate from the MoS₂ channel is analyzed. For this purpose, a two-dimensional electronic device consisting of hBN/MoS₂ is modeled, and identify candidate defects in hBN substrate that are likely to induce RTN through equilibrium condition. Subsequently, using MS-DFT simulations, we investigate the physical dynamics of these defects under varying temperatures and gate voltages. By employing configuration coordinate diagrams and a nonradiative multi-phonon model, we examine and discuss the unique characteristics of each defect.한국과학기술원 :전기및전자공학부
통계적 샤딩 및 텐서-트레인 분해를 활용한 추천 모델 가속 연산 스토리지 시스템
No full text학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[iv, 32 p. :]Deep Learning Recommendation models (DLRMs) play an essential role in providing personalized content in web applications such as social network services and video streaming platforms. However, as their performance improves, the parameter size of these models has reached terabyte (TB) scales, with memory bandwidth demands exceeding TB/s levels. Additionally, the computational heterogeneity of layers in the DLRM has highlighted the importance of configuring recommendation systems to match their computational characteristics. In this paper, we present SCRec, a scalable computational storage (CS) system with statistical sharding and tensor-train (TT) decomposition designed for recommendation models. SCRec employs a software framework featuring 3-level sharding that utilizes three different memory devices to effectively split hot and cold data, while balancing workloads. Additionally, SCRec implemented embedding and multi-layer perceptron (MLP) cores to accelerate computations in the embedding and MLP layers, improving throughput for the DLRM. By integrating the software framework and hardware accelerators with multiple SmartSSD devices, SCRec minimizes host communication overhead through peer-to-peer (P2P) data transfer. Our evaluation shows that SCRec demonstrates significant improvements in inference performance, achieving up to 55.77 gains compared to CPU-DRAM systems without accuracy degradation and up to 13.35 energy efficiency gains compared to multi-GPU systems.한국과학기술원 :전기및전자공학부
Enhancing thermal stability of hafnia ferroelectric thin film and transistor gate stack via functional interlayer insertion
No full text학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[iv, 35 p. :]현행 전하 트랩 방식의 3D 낸드 (NAND) 플래시 메모리는 고전압 동작을 기반으로 하고 있어, 집적도 향상을 위한 피치 크기 감소 시 따른 셀 간 간섭이 증가하는 등의 문제가 발생한다. 전하 트랩 방식은 집적도 향상의 한계에 도달할 것으로 예상되어, 저전압 동작이 가능한 차세대 3D 낸드 플래시 메모리 개발을 위해서 강유전체 소재가 연구되고 있다. 그 중 하프니아 (HfO2) 기반의 유전체 소재는 CMOS 공정과의 호환성 및 스케일링 가능성, 그리고 낮은 동작 전압만으로 분극의 반전이 가능한 성질 덕분에 활용 가능성을 더욱 주목받고 있다. 하프니아 강유전체 소재의 3D 낸드 플래시 메모리로의 상용화를 위해서는 박막의 열적 안정성이 고려되어야 한다. 이는 현행 3D 낸드 플래시 메모리 구조에서 강유전체 박막의 증착이 고온의 다결정질 실리콘 채널 증착 공정보다 선행되며, 채널의 온 전류를 증가시키기 위해 추가 열처리가 요구되기 때문이다. 따라서 3D 낸드 플래시 메모리에 적용되는 강유전체 박막은 높은 온도의 후속 열처리에 노출되는 것이 불가피하다. 그러나 하프니아 강유전체는 특유의 물성으로 인해 고온 및 장시간의 열처리에 노출되었을 때, 상유전성을 나타내는 결정상으로 전이하며 강유전성이 열화될 뿐만 아니라, 박막 내의 산소 공공의 증가로 누설 전류가 증가하는 등 열적 불안정성을 가진다.본 연구에서는 하프니아 강유전체 박막의 열적 안정성 향상을 위해 Hf0.5Zr0,5O2 (HZO) 박막 내 기능성 중간층을 삽입하는 방법을 제안한다. 중간층이 삽입된 Metal-Ferroelectric-Metal (MFM) 커패시터 구조에서는 800℃ 30분의 후속 열처리 인가 후에도 우수한 강유전성 (2Pr ≈ 24 μC/cm2) 및 내구성 (≈ 5.0 × 104 cycle) 을 확보할 수 있었다. 또한 하프니아 강유전체 전계 효과 트랜지스터 (Ferroelectric Field Effect Transistor: FeFET)의 게이트 스택으로서 연구되고 있는 Metal-Interlayer-Ferroelectric-Interlayer-Semiconductor (MIFIS) 구조에 중간층을 삽입하였을 경우, 800℃ 30분의 후속 열처리를 인가한 후에도 게이트 스택의 메모리 윈도우 특성 유지 및 프로그램/이레이즈 인가전압에 대한 안정성 향상 효과를 확인하였다. 본 연구를 통해 FeFET의 열적 안정성 향상을 위해서는 게이트 스택 내 중간층 삽입이 필요함을 확인하고, FeFET 게이트 스택의 열적 안정성에 대한 후속 연구들의 방향성을 제시하고자 하였다.한국과학기술원 :전기및전자공학부
수동형 광가입자망을 위한 간소화된 코히어런트 전송 시스템
No full text학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[iii, 53 p. :]Passive optical networks (PONs) have long regarded as a key technology for subscriber networks, gaining widespread adoption around the world. PON technology offers many advantages, including high reliability, low implementation cost, and ease of maintenance, making it an essential part of broadband access network. As the demand for higher data traffic continues to increase, the intensity-modulation and direct-detection (IM/DD) scheme commonly employed for PON systems face limitations in terms of bandwidth efficiency and the number of optical network units. To address these challenges, the coherent detection technique has been considered for future PON systems. Coherent PON systems offer the potential for enhanced sensitivity and full compensation of linear optical impairments, but accompany higher implementation and operational cost. In this thesis, I propose and demonstrate a simplified coherent transmission technology for very-high-speed PON systems, aiming to strike a balance between the cost and performance. For this purpose, I propose employing the polarization-division-multiplexed intensity modulation utilizing the electro-absorption modulator, instead of costly IQ modulator. Also, I propose adopting the heterodyne coherent detection to lower the insertion loss of the optical front-end and also to reduce the number of receivers and analog-to-digital converters at the receiver. The use of intensity modulation also obviates the need for the complicated carrier phase estimation typically required in the coherent receiver. I have carried out the proof-of-concept experimental demonstration using the 25-Gb/s OOK and 50-Gb/s PAM-4 signal. A single polarization is utilized in this demonstration. The results show that I can achieve the receiver sensitivities (@BER=10-2) of -34 and -18 dBm for the OOK and PAM-4 signals, respectively. The power budgets are estimated to be 40 and 24 dB if the transmitter power is 6 dBm. When the polarization-division multiplexing is applied to double the data rate, the power budget would be reduced by ~4 dB due to the polarization-beam splitter/combiner. The power budget for 50-Gb/s PAM-4 is short of the budget required in conventional PON systems. However, I believe that the performance could be enhanced by improving the DSP algorithm at the receiver and by increasing the transmitter power to 10 dBm. The findings of this thesis could be used to implement very-high-speed PON systems capable of providing 100-Gb/s service in a cost-effective manner.한국과학기술원 :전기및전자공학부