전산모사 및 I-V 모델링을 이용한 20-nm 노드 Ge2Sb2Te5 상변화 메모리 특성 분석

장현동 포항공과대학교 일반대학원 2020 국내석사

Phase-change memory (PCM) using chalcogenide materials has been spotlighted as non-volatile memory, which can fill the gap of memory hierarchy between DRAM and NAND Flash. However, as research on scaling down PCM devices has been actively conducted, problems such as resistance drift, cell-to-cell disturbance, undesired programming, and high RESET current have emerged. In particular, the chalcogenide material in the amorphous state has caused side effects, thereby raising the need for research for in nm-class PCM devices. Therefore, in this thesis, electrical and thermal performances of the PCM are characterized and discussed using technology computer-aided design (TCAD) and integrated circuit characterization and analysis program (IC-CAP) tools. Electrical and thermal performances of 20-nm node PCM described the bandgap model are extensively analysed according to physical parameters and geometry using fully-calibrated TCAD simulation. Increasing the maximal crystallization rate (r0) and decreasing the activation energy (Eact) reduces SET resistance and SET programming current, thus resistance ratio can be increased and consumes less power. This result suggests that Eact is more sensitive to electrical performance than r0. SET and RESET current decreases as impact ionization factor (II) increases. Also, programming current decreases and heat efficiency increases as thermal boundary resistance (TBR) thermal conductivity decreases and TBR metal resistivity increases. Decreasing the cell height reduces SET resistance, so increases read latency. However, current which produces the same Joule heating effect increases, thus increasing power consumption. Therefore, read latency and programming current are in a trade-off relationship when the cell height varies. Finally, these parameters can vary the threshold voltage, thus when designing the PCM cell, these parameters must be considered to meet the desired specifications. Threshold switching and snap-back mechanism is explained by the non-equilibrium carrier distribution and non-uniformity of electric field along the amorphous layer. Then, the appropriate barrier lowering change model that describe carrier transport in the subthreshold region is selected by comparing two models. Therefore, the subthreshold region, threshold switching, negative differential resistance, and ON region are implemented using analytical model of PCM. As a result, this model can be applied to circuit simulations of PCM devices with current equations based on physical computation.

An Analysis on the Consumption Structure of Contemporary Popular Culture with Network Science

이정우 포항공과대학교 융합대학원 2022 국내석사

Globalization and the development of information technology have enabled people in different societies to share their culture and consume cultural products through digital devices. This social change has made contemporary popular culture to transcend the borders between countries and penetrate the daily lives of consumers. Our thesis focused on investigating which social factors affect the consumption structure of contemporary popular culture. We constructed a consumption network of mobile games between countries to reflect the characteristics of contemporary popular culture. Using Hofstede's cultural dimensions theory and Facebook's social connectedness index, we revealed cultural distance and social ties between countries play important roles in shaping the consumption structure.

User Characteristics by News platform and Comment generation : Multiple methods approach for Interest and Interpretation of User

이민구 포항공과대학교 융합대학원 2023 국내석사

Depending on the feature of the media, the characteristics of the users are fixed. As a result, the information selection and interpretation may vary. Therefore, media study should focus on the process of user accepting and interpreting messages rather than simply looking at media as a tool for information delivery. In this study, Internet portal news and Online video platform news are compared from the audiences- centered perspective. Through comparison, we would like to explore whether user characteristics according to media platform appear in comments. For analysis, NAVER and YouTube, the representative of internet portal and Online video platform were analyzed. We collected the same news contents uploaded to each platform and built a dataset of 2,145,698 comments. As a result, it was found that comments of NAVER and YouTube not only concentrated on different news contents by reflecting the characteristics of users, but also there are differences in the trend of comments generation. Also, there are difference in the way how they interpreted the message of news contents. Through this, it was revealed that the characteristics of the user by media platform are actually impact on interest and interpretation of users.

A Study on the Improvement of Indicators for Clinical Decision Support Using Electronic Health Record Data

김혜영 포항공과대학교 융합대학원 2024 국내석사

본 연구는 전자의무기록 데이터와 머신러닝 모델을 활용하여 의료진의 임상 의사결정을 보조하는 지표를 개선시키기 위한 방법론을 제시하는 것을 목적으로 한다. 이를 위해, 중환자실에서의 충분한 의료 처치 여부를 기준으로 환자의 중환자실 퇴실 가능 여부를 판단하는 환자 상태 평가 지표의 개선을 사례 연구로 활용하였다. 전자의무기록 데이터와 머신러닝 모델을 활용한 지표 개선 과정은 다음과 같다. 첫째, 개선할 지표의 목적에 부합하는 예측 문제를 설정한다. 둘째, 해당 문제의 결과를 예측하기 위한 머신러닝 모델을 학습시킨다. 셋째, 모델에 활용된 중요 변수를 분석하고, partial dependence를 활용하여 지표 항목을 생성한다. 마지막으로, 모델의 중요 변수들과 기존 임상 현장에서 사용되던 지표의 항목들을 비교 분석하여 지표의 개선점을 탐색한다. 본 연구는 중환자실 퇴실 판단을 위한 환자 상태 평가 지표의 개선을 예시로 위에서 제시된 방법론의 결과를 확인하고자 하였다. 먼저, 중환자실 재입실 시점 간격에 따른 로지스틱 회귀 모델과 오즈비의 비교를 통해, 중환자실 퇴실 후 48시간 이내의 재입실이 중환자실의 의료 서비스가 환자에 미치는 영향을 잘 고려하는 재입실 정의임을 확인하였다. 이어서, 중환자실 퇴실 이후 48시간 이내의 재입실을 예측하는 머신러닝 모델들을 개발하고, 이 모델들의 성능이 베이스라인으로 볼 수 있는 기존의 퇴실 판단 지표보다 우수하게 나타내는 것을 확인하였다. 로지스틱 회귀, 랜덤 포레스트, 나이브 베이즈, 그레디언트 부스팅 등의 머신러닝 모델들 중 그레디언트 부스팅 모델이 가장 높은 예측 성능을 보였다. 마지막으로, 그레디언트 부스팅 모델을 활용하여 중환자실 환자들의 퇴실 판단에 유용한 중요 변수들을 추출하고, partial dependence 분석을 통해 이러한 변수들의 영향을 검토하였다. 이러한 변수들을 기반으로 한 퇴실 판단 지표의 예측은 기존 지표를 통한 예측보다 개선된 성능을 보였다. 또한 기존 퇴실 판단 지표의 변수들과는 달리, 환자 상태 경고 알람 횟수와 같은 새로운 변수들이 중요 변수로 도출되었으며, 이는 지표의 보완에 기여할 수 있음을 확인하였다. 본 연구에서 활용한 임상 의사결정 보조 지표의 개선 방법론은 병원 내의 다양한 임상 의사결정 보조 지표의 개선에 활용되어, 병원 내 의료서비스 관리에 기여할 수 있을 것이다. 복잡한 모델이 끊임없이 개선되고 활용될 수 없는 병원 환경에서도 이러한 방법론을 통한 지표의 개선은 의료진이 보다 정확하고 신속한 의사결정을 내리는 데 도움이 될 것으로 기대된다. This study aims to present a method to improve indicators that support medical staff in clinical decision-making, by using Electronic Health Record (EHR) data and machine-learning (ML) models. Specifically, it focuses on improving patient status assessment indicators that determine the suitability of intensive care unit (ICU) discharge by quantifying the sufficiency of medical treatments provided in the ICU. The process of improving the indicators by using EHR data and ML involves four steps: (1) define a predictive problem that aligns with the purpose of the indicator to be improved; (2) build an ML model to predict outcomes for this defined problem; (3) interpret the important variables of the model and create indicator items by applying partial dependence analysis; (4) identify potential improvements by comparing the model’s significant variables to the existing clinical indicators. As an example of the proposed method, this study focused on improving patient status assessment indicators for ICU discharge decision-making. Initially, a comparison of logistic regression models and odds ratios that consider the interval of ICU readmission, confirmed that readmission within 48 hours after ICU discharge is a valid criterion to evaluate the effectiveness of ICU medical services on patients. Subsequently, ML models were developed to predict readmissions within 48 hours post-ICU discharge; these models were more accurate than the existing discharge-decision indicators, which can be considered as the baseline. Among various ML models, the Gradient Boosting model was had the highest predictive accuracy. Finally, using the Gradient Boosting model, important variables useful for ICU patient discharge decisions were extracted, and their effects were assessed using partial-dependence analysis. The discharge decision indicators that consider these variables showed higher accuracy than the existing indicators. Additionally, new variables (e.g., frequency of patient-status warning alarms) that differed from those in the existing discharge decision indicators, were identified as important; this result confirms their potential contribution to improving the indicators. The clinical decision-support indicator improvement method that is developed in this study can be applied to enhance various clinical decision support indicators within hospitals, contributing to better management of hospital medical services. Even in hospital environments in which complex models cannot be continuously improved and utilized, the improvement of indicators by using this method is expected to aid medical professionals to increase the speed and accuracy and of their decisions.

Nowcasting Korean GDP growth using Machine Learning with Economic Policy Uncertainty feature

정승민 포항공과대학교 융합대학원 2022 국내석사

GDP growth is an indicator of a country's economic situation and is a crucial factor in financial decisions. Nevertheless, since it has a problem of being announced lately, 'Nowcasting', the prediction of GDP growth at present, is being treated as an essential issue. Due to the recent increase in uncertainty, studies to increase the accuracy of Nowcasting are primarily divided into two directions. One is to reflect uncertainty as a variable, and the other direction is to use ML models as predictive models. However, there has yet to be an attempt to incorporate both approaches. Therefore, this study aims to integrate both approaches to generate a prediction model for the GDP of Korea. The proposed method first extracts common factors through the Dynamic Factor Model to reduce the dimensions of 83 economic indicators affecting GDP growth. Then, the Economic Policy Uncertainty value, an indicator of uncertainty, is combined with the reduced factors, and they are used as input features of prediction models. Finally, several machine learning models are used to predict GDPs. To validate the proposed approach, we conduct experiments with Korean GDP-related data. In the experiment, we construct two data sets with and without the Economic Policy Uncertainty value to explore the impact of the uncertainty. Random Forest, Gradient Boost, and XGBoost are used as ML-based prediction models, while OLS regression is used as a conventional prediction model. The experimental result shows that including the EPU feature provides higher prediction accuracies for all four models. In addition, the performances of the ML models are more elevated than that of OSL regression.

The design of all-dielectric metasurface for selectively blocking Near-Infrared region of solar spectrum

김은종 포항공과대학교 일반대학원 2020 국내석사

Blocking near-infrared region (NIR) is indispensable for saving energy consumed to maintain an interior temperature in buildings. The methods of blocking NIR are divided into two types. One is blocking NIR without visible light for applying to windows. The other is blocking the overall solar spectrum including visible light because the visible light accounts for 45% in the solar spectrum. Thus, we designed the metasurface of two types, respectively. In the transparent metasurface blocking NIR, simultaneously enhancing transmission in visible light and blocking in NIR remains challenging. Here, we demonstrate the trans-parent all-dielectric metasurface selectively blocking the NIR by using TiO2 nanocylinder and ITO layer. The ITO layer is implemented as a back reflector because the ITO is trans-parent in visible light whereas the ITO becomes reflective materials in the long-wavelength region (λ > 1500 nm). The designed metasurface exhibits a high average transmittance of 70% in visible light and high solar energy rejection (SER) of 90% in NIR. Furthermore, the performance of the designed metasurface is maintained over a wide range of an incident angle of light. Therefore, the metasurface gives an advanced guide-line for design energy-saving applications. In the opaque metasurface blocking NIR, the previous all-dielectric metasurfaces have difficulty in reflecting overall NIR because the reflection region is too narrow. Here, we demonstrate the all-dielectric metasurface blocking almost overall NIR by using amor-phous Si (a-Si) and SiO2. Since amorphous Si has a high refractive index (~3.3) and high extinction coefficient in NIR, the designed metasurface exhibits high reflection (1050 nm ≤ λ ≤ 2320 nm) as well as high absorption (λ < 1040 nm), thereby leading to high solar energy rejection (SER) of 94% in NIR. The performance of the designed metasurface is independent over a wide range of incident light. Furthermore, a-Si substrate constituting the metasurface can be readily deposited on other materials such as glass and plastic film, so the proposed metasurface has high applicability for a large-area fabrication rather than crystalline Si and GaAs.

Design and Fabrication of Nanoscale Metal Interconnections for Transparent Deformable Electronic Devices

김동욱 포항공과대학교 일반대학원 2021 국내박사

As transparent displays and touch screens begin to be introduced to the public, the technology for transparent deformable electronic devices is attracting enormous attention as a next-generation electronic technology. Deformable optoelectronic devices, such as displays, solar cells, touch screens, and smart windows that maintain their functions under mechanical deformations have been developed, and various approaches to transparent deformable electrodes have been studied intensively. Despite these interests, transparent deformable nano- and micro-scale integrated interconnections that are easy to be patterned and positioned are receiving somewhat less attention. Since the most successful and feasible concept leading to deformable devices is linking rigid islands of active device components (transistor, light-emitting devices, photovoltaics, etc.) with deformable interconnections, developing transparent interconnections that can retain good electrical performance under high mechanical strain is highly required. In this work, I designed and fabricated three different types of transparent deformable nanoscale metal interconnections; (i) One-dimensional (1D) metal nanolines which were deposited on the flexible substrates by simple and reliable nanofiber (NF) photolithography, (ii) 1D wavy stretchable single metal NF which were individually positioned by the electrohydrodynamic (EHD) printing and metallized through room-temperature electroless plating, and (iii) Two-dimensional (2D) foldable and stretchable gold (Au) film hybrid electrodes which were composed of the anisotropic conductive ultrathin films (ACUFs) and the ultrathin Au film electrodes. These fabricated metal interconnections were not only electrically deformable but also optically transparent due to their nanoscale dimensions, and were able to be individually positioned and patterned in desired positions, shapes, and alignments. Also, all three deformable interconnections are fabricated by low-temperature processes and can easily expanded to large-scale production. The fabricated nanoscale metal interconnections were used as the interconnecting electrodes in the transparent and deformable field-effect transistors (FETs) array and the transparent electrodes of the deformable light-emitting devices.

Measuring Democratic Values Oriented by AI News Recommendation Algorithms

황수현 포항공과대학교 융합대학원 2023 국내석사

The market for news in Korea is increasingly using AI. The Korea Press Foundation estimated that 79.2% of Koreans would view news online portal sites in 2021, and Naver and Kakao had already transitioned from human to AI editing of news recommendations in 2017, accounting for more than 90% of the market share for portals. The truth, however, is that there is no established yardstick for judging AI news, and this is accompanied by a lack of awareness of AI’s market share in the news. The purpose of this study was to evaluate AI news recommendations in light of the democratic values that Korean society pursues. We propose three models of democracy, liberal, participatory, and deliberative, to achieve this purpose, in addition to evaluating how well Naver’s and Kakao’s AI-recommended news adhered to the principles of each democracy model, i.e., freedom, social involvement, and discussion. We identified that Kakao is closer to the deliberative democracy model, whereas Naver is closer to the liberal model, such that Naver’s news recommendation algorithm is better suited to guaranteeing people’s freedom of news selection. On the other side, we interpret that under Kakao’s algorithm, members are contributing substantially more to attaining agreement via debate. All democratic values are relative, and we cannot discern between excellent and terrible algorithms merely because they concentrate on a certain value. The key is to understand where algorithms fit into the map of values that each culture pursues, and this study is an effort to do just that. This is also crucial because it sets a precedent for the ultimate objective of putting in place an algorithm that complies with Korean society’s values. 한국 뉴스 시장에서 AI가 차지하는 비중은 갈수록 증가하고 있다. 한국언론진흥재단에 따르면 2021년 한국인의 79.2%는 뉴스를 볼 때 인터넷 포털 사이트를 이용하며, 포털 시장 점유율 90% 이상을 차지하는 네이버와 카카오는 2017년에 이미 뉴스 추천을 사람 편집에서 AI 편집으로 전환했다. 그러나 뉴스 시장에서 AI가 차지하는 비중에 대한 인식이 부족할뿐더러, AI 뉴스를 평가하는 합의된 기준 또한 부재한 것이 현실이다. 이 연구는 AI가 추천한 뉴스를 우리 사회가 추구하는 민주적 가치에 기준하여 평가하기 위한 것이다. 이를 위해 3가지 민주주의 모델 – 자유주의, 참여주의, 심의주의 – 을 상정하고, 네이버와 카카오의 AI 추천 뉴스가 각 민주주의 모델이 추구하는 가치 – 자유, 사회 참여, 토론 및 합의 – 와 얼마나 가까운지, 혹은 먼지를 측정하였다. 실험 결과 네이버는 자유주의 모델에, 카카오는 참여주의와 심의주의 모델에 더 가까운 것으로 나타났다. 이는 네이버 뉴스 추천 알고리즘은 개인이 뉴스를 선택할 자유를 보장하는 데 비교적 더 많이 기여하는 반면, 카카오 뉴스 추천 알고리즘은 구성원들이 사회의 다양한 사안에 참여하고, 토론하여 합의를 이루는 데 상대적으로 더 많이 기여하고 있다는 의미로 해석할 수 있다. 민주적 가치는 모두 상대적이며, 특정 가치에 집중한다는 이유로 좋은 알고리즘과 나쁜 알고리즘을 가를 수 없다. 중요한 것은 각 사회가 추구하는 가치의 지형도에 알고리즘이 어떻게 위치하고 있는지를 인지하는 것이며, 이 연구는 그 인지를 위한 작업이다. 이는 또한 한국 사회의 가치에 부합하는 알고리즘 구현이라는 최종 목표의 선행 작업이라는 데 의의가 있다.

Automatic Attribute Extraction using Domain Expertise in Fashion and its Evaluation on Compatibility and Diagnosis

배승예 포항공과대학교 융합대학원 2022 국내석사

Predicting outfit compatibility refers to determining whether fashion items look good if worn together. In recent years, a few studies have used visual-semantic space in outfit compatibility prediction and proposed modeling or equations to capture high-level information. However, the proper textual attributes help to form a more accurate visual-semantic space, and providing the domain-specific details allows the compatibility model to learn semantically more robust information. This thesis proposes a method to extract the domain-specific fashion attributes using color expertise. The proposed method maps the pattern and adjectives corresponding to the closest one among Kobayashi’s color triplets of each item as fashion style concepts. Then, it adjusts the resulting concepts by zero-shot classification of fine-tuned CLIP to make them more distinguishable. Experiments with four datasets that differ in the composition of the extracted concepts in text attributes are conducted to validate the proposed method. The dataset, including adjusted fashion style concepts, outperforms the prior baseline with a 14% increase in FITB accuracy of outfit compatibility prediction. The result shows that high-level semantic features are prominent to get the unified representation through visual-semantic space and verifies that our approach is more applicable to the fashion domain for outfit diagnosis.

Instagram Post Lifestyle Classification Model for Effective Influencer Marketing

한유정 포항공과대학교 융합대학원 2024 국내석사

With the advent of social media, the 'influencer' has emerged. An influencer is an individual with a significant number of followers on their personal social media account. However, from a corporate perspective, finding multiple influencers suitable for their brand and products remains challenging. Most research into the compatibility between brands and influencers has been predominantly confined to methods such as surveys and case studies to verify effectiveness. In other words, there is a notable absence of technical research focused on identifying the 'aesthetic harmony' between brands and influencers, which is a vital component in influencer marketing. Therefore, this study proposes a scenario-based application case, which involves developing a model that classifies Instagram post lifestyles, using a fine-tuned CLIP model with crawled data. Specifically, for the task of lifestyle classification, we crawled images and texts that can describe five lifestyles using seventy-one adjectives as keywords based on the word-image scale. Secondly, for efficient natural language supervised learning during CLIP fine-tuning, text preprocessing is performed. Ultimately, we demonstrate that our proposed model achieves high performance with 87% accuracy compared to four baseline models. Furthermore, we proposed a process for ranking influencers suitable for brands using the style model, thereby providing a practical guide for its application.