二維材料的合成掃描穿隧顯微影像：深度學習分析流程

Abstract

掃描式穿隧顯微鏡（STM）是一種極為有效的技術，可用於成像導電或半導體材料表面的影像，其解析度可達單一原子。然而，STM 仍受限於對人力密集的專家解讀的依賴，以及容易受到探針狀況、掃描參數與硬體相關失真所引入的人為誤差影響。這些因素限制了其速度與可擴展性。

在本研究中，我提出了一套以物理建模為基礎的模擬框架，用於生成合成的 STM 影像。該系統利用緊束縛方法計算電子結構，建構真實的探針幾何形狀，並模擬由 PID 控制迴路與壓電致動器行為所引起的失真。這使得能夠生成反映 STM 成像實際複雜性的龐大標註資料集，為訓練機器學習模型提供基礎。

利用這些合成資料集，我訓練了一個用於鎢二硒化物缺陷定位的深度學習模型，並將其表現與未考慮物理失真的模型進行比較。對真實 STM 測量結果的定性評估顯示，完整模擬所訓練出的模型具有更佳的泛化能力，證明高擬真模擬對於推進 STM 影像分析自動化的潛力。

Scanning Tunneling Microscopy (STM) is a highly effective technique for imaging the surfaces of conductive or semiconductive materials with a resolution down to single atoms, but it remains limited by its reliance on labor-intensive expert interpretation and susceptibility to artifacts introduced by tip condition, scanning parameters, and hardware-related distortions. These factors hinder its speed and scalability.

In this work, I present a simulation framework for generating synthetic STM images grounded in physical modelling. The system calculates electronic structure using the tight-binding method, models realistic tip geometries, and simulates distortions arising from PID control loop and piezoelectric actuator behaviour. This allows for the generation of large labelled datasets that reflect the real-world complexity of STM imaging, providing a foundation for training machine learning models.

Using these synthetic datasets, I trained a deep learning model for defect localization in tungsten diselenide, comparing its performance against a model trained without physical distortions. Qualitative evaluation on real STM measurements reveals superior generalization when using the full simulation, demonstrating the potential of high-fidelity simulations to advance automated STM image analysis.