Deep Learning-based Automated Crystallographic Analysis of HfO2 Thin Films : High-Throughput STEM Image Analysis = 딥러닝 기반 HfO2 박막의 자동 결정학적 분석: 고속 STEM 이미지 분석

다국어 초록 (Multilingual Abstract)

This study aims to develop a deep learning–based framework for automated quantitative analysis of crystallographic properties (phase, zone axis, and growth direction) in Y-doped HfO₂ epitaxial thin films using atomic-resolution HAADF-STEM images. Because the ferroelectric properties of HfO₂ are strongly governed by the volume fraction of the metastable orthorhombic (O) phase and the crystallographic growth direction, quantitative evaluation of these parameters is essential. However, in epitaxial thin films, interfacial structural complexity—including gradual phase transitions—introduces substantial challenges for quantitative analysis using conventional approaches such as XRD, 4D-STEM, and manual indexing of HAADF-STEM images.
Accordingly, this research was conducted by (i) constructing a deep learning model that classifies 100×100-pixel atomic-resolution HAADF-STEM image patches into 11 classes representing combinations of phase, zone axis, and growth direction, and (ii) developing an evaluation framework that extends these patch-level predictions to pixel-wise crystallographic information in full-size images. To ensure model robustness, experimental and simulated datasets were combined and augmented using diverse image augmentation techniques. For full-size image analysis, structure maps and uncertainty maps were generated via window-based sampling and a cascading ensemble strategy, while scale sensitivity and computational efficiency were simultaneously improved through a multi-scale ensemble and an efficient window-sampling scheme.
When applied to Y-doped HfO₂ epitaxial thin films, the proposed framework enabled pixel-wise quantification of phase fraction as well as distributions of zone axis and growth direction over large areas. In addition, the uncertainty map effectively visualized structural ambiguity in regions exhibiting phase evolution, such as phase/domain boundaries and the substrate–film interface. Furthermore, quantitative analysis across films of different thicknesses demonstrated the performance of the framework by statistically comparing epitaxial-strain-induced phase stabilization and thickness-dependent phase-transition behavior.

국문 초록 (Abstract)

본 연구는 HAADF-STEM 원자해상도 이미지로부터 Y-doped HfO₂ 에피택셜 박막의 결정학적 특성(상, zone axis, 성장방향)을 자동으로 정량 분석하기 위한 딥러닝 기반 프레임워크 개발을 목표로 한다. HfO₂ 강유전 특성은 준안정 orthorhombic(O)상의 분율과 결정학적 성장방향에 크게 좌우되기 때문에 이를 정량적으로 분석하는 것이 중요하다. 반면 에피택셜 박막에서는 점진적인 상전이를 포함한 계면 구조 복잡성으로 인해 XRD, 4D-STEM, HAADF-STEM 수동 인덱싱 등의 방법론들은 이를 정량적으로 분석하는데 일정한 한계가 있다.
이에 본 연구에서는 (i)원자해상도 HAADF-STEM 패치 이미지(100×100)로부터 상/zone axis/성장방향을 11개 클래스로 분류하는 딥러닝 모델을 구축하고, (ii)해당 예측을 평가 이미지의 픽셀 단위 결정학 정보로 확장·정량화하는 전략으로 프레임워크를 개발하였다. 이를 위해 실험·시뮬레이션 데이터를 결합하고 다양한 이미지 증강기법을 활용하여 모델의 강건성을 확보하였으며, 이미지 분석의 경우 윈도우 샘플링과 케스케이딩 앙상블을 통해 구조맵과 불확실성맵을 도출하고, 다중 스케일 앙상블과 윈도우 샘플링 기법을 통해 배율 민감도와 계산 효율 문제를 동시에 완화하였다.
관련 결과에 따르면, 대면적의 Y-doped HfO₂ 상분율, zone axis 및 성장방향 분포를 픽셀 단위로 정량화할 수 있었으며, phase/domain boundary 및 기판-박막 계면과 같이 상변화가 나타나는 영역에서 불확실성맵을 통해 구조적 불확실성을 효과적으로 시각화할 수 있었다. 또한 두께가 다른 박막 시편에 대한 정량 분석을 통해 상 안정화 및 두께 의존적 상전이 거동을 통계적으로 비교·평가하며 개발된 분석 프레임워크의 성능을 검증하였다.

목차 (Table of Contents)

• Chapter 1. Introduction 1
• 1.1. Background and motivation 1
• 1.2 Data-driven approaches in materials science 3
• 1.3 Problem statement and research objectives 6
• Chapter 2. Deep learning framework for automated crystallographic phase identification 9
• 2.1. Feasibility study 9
• 2.1.1. Proposed methodology and workflow 9
• 2.1.2. Feasibility validation with simple CNN 12
• 2.2. Dataset construction and augmentation 19
• 2.2.1. Physics-based augmentation 22
• 2.2.2. Conventional augmentation 25
• 2.3. Transfer learning model development 27
• 2.4. Model performance and validation 30
• 2.4.1. Overall classification accuracy 30
• 2.4.2. Detailed performance of optimized model 32
• 2.4.3. Error analysis and model validation 35
• 2.5. Conclusion 39
• Chapter 3. Quantitative materials characterization using structure and entropy maps 41
• 3.1. Overview of full-size image analysis framework 41
• 3.2. Window-based ensemble 43
• 3.2.1. Voting ensemble for structure map 43
• 3.2.2. Average ensemble for entropy map 45
• 3.3. Multi-scale ensemble 45
• 3.4. Random window sampling 48
• 3.4.1 Limitation of sliding window sampling 50
• 3.4.2 Random window sampling with constraints 50
• 3.5. Application to Y-doped HfO₂ thin films 51
• 3.5.1. Structure map and quantitative analysis 51
• 3.5.2 Entropy map: substrate-film interface and structural disorder analysis 56
• 3.5.3 Thickness-dependent phase evolution and strain effect 59
• 3.6 Conclusion 62
• Chapter 4. Summary 64
• 4.1. Development and validation of the deep learning model 64
• 4.2. Development of the quantitative analysis framework 65
• Conclusion 66
• Reference 67
• 국문초록 (Abstract in Korean) 72