Inversion of dynamical electron scattering for atomically-resolved structural analysis

4D扫描透射电子显微镜（4D- stem）方法包括扫描电子探针穿过样品，并在每个光栅位置记录衍射图案。通过迭代算法，可以重建原子尺度的结构和化学。在这个项目中，现代纳米结构被用来展示结构表征的精度提高到皮米范围，并最大限度地提高化学灵敏度，从而探索目前隐藏的结构-性质关系。通过将冻结声子近似中的散射问题表述为一个神经网络，消除了热扩散散射（TDS）的持续模型违反。这种网络是基于实验训练的，通过对任意单物理参数的梯度的有效计算，得到原子位置和类型作为直接测量结果。该项目有四个主要目标：1。BaTiO3中的铁电畴是量化的，而O-和ti -离子与其对称位置的差异仅为几微米。由于具有1-2nm厚层的隧道结在未来的数据存储设备中很有前景，因此对应变诱导畴的pm级映射应该能够揭示纳米层系统中畴形成的物理学以及与观察到的铁电畴不可切换性有关的器件故障。2. 在原子尺度上定量测量轴向InAs/AlAsSb纳米线异质结构中的偏析，从而了解高电子迁移率晶体管（HEMT）的电子性能。通过明确地利用TDS，我们实现了与已建立的平面图相比的增强化学灵敏度，取代了Z-contrast STEM中固定的前向散射模型，并消除了界面处复杂的动态散射。3. 研究了合金二硫化物单层Mo{S(x)Se(1-x)]的化学结构。合金二维晶体使一个新的自由度x调整电子带对准在异质层。关于组vi位的化学组成的梯度被用来以范例的方式设计实验条件。低剂量4D-STEM实验用于绘制vi族单体、同聚体和异源二聚体，并能够理解可见光获得的光谱。4. 对重建算法进行了扩展，使得化学键可以在实验中获得。仿真和实验用于开发附加网络层并建立其参数化。通过实现ResNet模块，研究了网络对试件厚度的可微性。综上所述，本文开发并发布了一种通用的重建方法，供社会科学使用。

The method of 4D scanning transmission electron microscopy (4D-STEM) consists of scanning an electron probe across a specimen, and recording a diffraction pattern at each raster position. By iterative algorithms, atomic-scale structure and chemistry can be reconstructed. In this project, contemporary nanostructures are used to demonstrate precision enhancement of the structural characterization to the picometre-range, and to maximise chemical sensitivity so as to explore currently hidden structure-property relationships. By formulation of the scattering problem in frozen phonon approximation as a neural network, persistent model violations as to thermal diffuse scattering (TDS) are eliminated. Such networks are trained based on experiments via efficient calculation of gradients with respect to arbitrary solely-physical parameters, yielding atom positions and types as the direct measurement result. The project addresses four main objectives: 1. Ferroelectric domains in BaTiO3 are quantified, whereas O- and Ti-ions differ from their symmetry positions by a few picometres only. Since tunnel junctions with 1-2nm thick layers are promising for future data storage devices, pm-level mapping the strain-induced domains is supposed to shed light on the physics of domain formation in nanolayer systems and device failure regarding the observed non-switchability of ferroelectric domains. 2. Segregation in axial InAs/AlAsSb nanowire heterostructures is measured quantitatively at the atomic scale, so as to understand the electronic performance of high electron mobility transistors (HEMT). By exploiting TDS explicitly, we achieve enhanced chemical sensitivity in comparison to established ptychography, the replacement of fixed forward scattering models in Z-contrast STEM, and the elimination of the complex dynamical scattering at interfaces. 3. Alloyed dichalcogenide monolayers Mo{S(x)Se(1-x)] are explored chemically and structurally. Alloyed 2D crystals enable a new degree of freedom x to tune the electronic band alignment in heterobilayers. Gradients with respect to the chemical composition of group-VI sites are used to design experimental conditions in paradigmatic manner. Low-dose 4D-STEM experiments serve to map group-VI monomers, homomers and heterodimers and enable the understanding of spectra obtained with visible light. 4. The reconstruction algorithm is expanded such that chemical bonding becomes experimentally accessible. Simulations and Experiments are used to develop additional network layers and setting up their parametrization. By implementing ResNet modules, differentiability of the network as to the specimen thickness is studied. In summary, a generic reconstruction method is developed and released for scientific use in the community.