← 논문 노트 목록
AI

Unsupervised deep denoising for four-dimensional scanning transmission electron microscopy

Alireza Sadri(제1저자), Timothy C. Petersen, Emmanuel W. C. Terzoudis-Lumsden, Bryan D. Esser, Joanne Etheridge, Scott D. Findlay(교신저자)

npj Computational Materials (2024) · DOI: 10.1038/s41524-024-01428-x · 넣은 날 2026-08-09

1. 서지정보

Sadri, A., Petersen, T. C., Terzoudis-Lumsden, E. W. C., Esser, B. D., Etheridge, J., & Findlay, S. D. (2024). Unsupervised deep denoising for four-dimensional scanning transmission electron microscopy. npj Computational Materials, 10, 243. https://doi.org/10.1038/s41524-024-01428-x

저자 메모: 주니가 참여한 논문이 아니다. SHINE(TEM 이미지 디노이징) 논문과 같은 결로, 이번엔 TEM 이미지가 아니라 4D-STEM(4차원 주사투과전자현미경, 각 탐침 위치마다 2차원 회절 패턴을 함께 기록하는 기법) 데이터를 딥러닝으로 디노이징하는 외부 논문이다.

2. 연구문제

4D-STEM은 시료를 스캔하는 각 탐침(probe) 위치마다 CBED(수렴빔 전자회절, convergent beam electron diffraction) 패턴을 하나씩 기록해서, 위치 정보(2차원)와 회절 패턴 정보(2차원)를 합쳐 4차원 데이터셋을 만드는 기법이다. 문제는 이 데이터가 검출기의 수많은 픽셀에 도즈(dose, 시료에 조사하는 전자량)를 분산시키다 보니 픽셀당 신호대잡음비가 떨어져 노이즈가 심하다는 것 — 특히 손상에 약한 시료는 도즈를 더 낮춰야 해서 노이즈가 더 심해진다. 기존 2D STEM 이미지용 디노이징 기법은 탐침 위치와 회절 좌표 사이의 연결 관계를 활용하도록 설계되지 않았다. 정답(clean) 데이터 없이, 노이즈 낀 4D-STEM 데이터 자체만으로 학습해 디노이징할 수 있는가?

3. 방법

4. 데이터

5. 결과

6. 한계

저자가 밝힌 한계:

읽으면서 느낀 한계:

7. 원문 인용

By simultaneously achieving high spatial and angular sampling resolution, four dimensional scanning transmission electron microscopy (4D STEM) is enabling analysis techniques that provide great insight into the atomic structure of materials. — Abstract

We demonstrate an unsupervised deep learning model that leverages the continuity and coupling between the probe position and the electron scattering distribution to denoise 4D STEM data. — Abstract

By restricting the network complexity it can learn the geometric flow present but not the noise. — Abstract

The need to train afresh on every new data set is the price of the unsupervised approach. — Discussion (Applications 절 말미, 한계 논의)

An assumption made here is that the raster scan is evenly spaced. — Discussion (한계 논의)

관련 개념

왜 TEM 관찰 개념과 다른가: 이 논문이 실제로 다루는 대상은 TEM(투과전자현미경)이 실공간에서 찍는 이미지가 아니라, STEM(주사투과전자현미경)이 각 탐침 위치에서 기록하는 CBED(수렴빔 전자회절) 패턴 — 즉 역공간(reciprocal space) 회절 데이터다. 4D-STEM은 STEM의 확장 기법이라는 점에서 넓게 보면 전자현미경 계열이 맞지만, “시료를 찍은 실공간 상(image)을 본다”는 TEM 관찰(투과전자현미경) 개념과는 다루는 데이터의 성격 자체가 다르다(회절 패턴 vs. 실공간 이미지). 그래서 이번 노트에서는 tem-observation 개념 페이지를 걸지 않았다.

[[4d-stem]] · [[geometric-flow-denoising]] · [[ptychography]] · [[cbed-pattern]]

(모두 이 지식체계에 아직 없는 개념이라 SCHEMA.md 규칙대로 “나중에 쓸 것” 표시로 남겨둔다.)