高分辨透射电子显微镜是在原子尺度上研究材料微观结构的重要工具。精确获取原子位置坐标是进一步实现材料结构定量分析的基础。由于电子显微镜中的图像噪音以及不可避免的存在各种像散，尤其是在原位透射电子显微镜实验中对高速成像需求，使得快速、高精度的获取原子坐标成为一大挑战。近年来，随着人工智能在电镜领域的深入应用，基于深度学习的电子显微图像分析也正逐步发展。基于以上背景，本文借助球差电子显微镜的积分差分相位衬度（iDPC）和高角环形暗场像（HAADF）以及基于深度学习模型的原子定位方法，研究了Co3O4催化剂中的表面相变和石墨烯基Fe-N-C单原子催化剂的活性位点配位结构。论文首先研究了利用深度学习模型确定高分辨透射电子显微图像原子柱坐标的方法，利用数据增强的模拟图像数据集训练了DeepLabV3+模型并应用于原子图像的识别。相比于传统二维高斯拟合方法，深度学习方法具有更好的鲁棒性，对图像中的噪音不敏感，在原子定位方面和二维高斯拟合方法有相近的精度。通过对连续多帧原子图像的识别，本文观察到了电子束诱导的Co3O4催化剂中的可逆表面相变。通过对表面相变过程中原子位置的动态精确分析，建立了表面相变的动力学过程，对理解其催化性质提供了结构基础。在上述工作基础上，论文进一步研究了燃料电池Fe-N-C非贵金属单原子催化剂的配位原子结构。目前，该催化剂的活性位点仍存在争议，而基于同步辐射和穆斯堡尔谱等表征方法得到的是平均的配位结构，无法识别实际配位结构的多样性。基于深度学习模型的HAADF和iDPC图像分析，论文在原子尺度上实现了对Fe单原子配位结构的精细表征，发现石墨烯基Fe-N-C催化剂的真实活性位点主要是位于边缘和台阶处的单原子Fe-N结构，催化剂中边缘（包括台阶）和面内的活性位点比例大致为3：1。对于面内Fe-Nx活性位点，周围碳六元环结构通常存在缺陷，造成Fe-N键长呈现非对称结构，不同于传统认识的Fe-N4-C平面对称结构，这种非对称的Fe-N4键长结构对其催化性质和稳定性将有显著的影响。通过密度泛函理论计算进一步研究了台阶处的Fe-N2位点对燃料电池氧还原反应的影响。论文所解析得到的原子尺度的配位结构信息将有助于深入理解Fe-N-C单原子催化剂的催化活性位结构和催化性质。

High-resolution transmission electron microscopy is an important tool for studying the microstructure of materials at the atomic scale. Accurately obtaining atomic position coordinates is the basis for further quantitative analysis of material structures. Due to the image noise in electron microscopy and the inevitable existence of various astigmatisms, especially the high-speed imaging requirements in in situ transmission electron microscopy experiments, it is a big challenge to obtain atomic coordinates quickly and with high precision. In recent years, with the in-depth application of artificial intelligence in the field of electron microscopy, electron microscopic image analysis based on deep learning is also gradually developing. Based on the above background, this thesis studies the surface phase transition in Co3O4 catalysts and the coordination structure of Fe-N-C single-atom catalysts by means of integral differential phase contrast and high-angle annular dark field images of spherical aberration electron microscopy and an atom positioning method based on deep learning models.The thesis first uses deep learning to determine the atomic column coordinates of high-resolution transmission electron microscopy images, and uses the data-enhanced simulated image dataset to train the DeepLabV3+ model and apply it to the recognition of atomic images. Compared with the traditional two-dimensional Gaussian fitting method, the deep learning method has better robustness, is not sensitive to the noise in the image, and has similar accuracy to the two-dimensional Gaussian fitting method in terms of atom positioning. Through the identification of continuous multi-frame atomic images, Co3O4 is a class of catalysts that are expected to be applied in the field of electrolysis of water for hydrogen production. In this paper, an electron beam-induced reversible surface phase transition in Co3O4 catalysts was observed. Through the precise analysis of the dynamics of atomic positions during the surface phase transition process, the kinetic process of the surface phase transition is established, which provides a structural basis for understanding its catalytic properties.On the basis of the above work, the paper further studies the coordination atom structure of Fe-N-C non-noble metal single-atom catalysts for fuel cells. At present, the active site of the catalyst is still controversial, and the average coordination structure is obtained mainly based on synchrotron radiation and M?ssbauer spectroscopy, which cannot identify the diversity of the actual coordination structure. Based on the HAADF and iDPC image analysis of the deep learning model, the paper realized the fine characterization of the single-atom coordination structure of Fe at the atomic scale, and found that the real active sites of graphene-based Fe-N-C catalysts are mainly located at the edges and steps In the single-atom Fe-N structure, the ratio of active sites on the edge (including steps) and in-plane in the catalyst is roughly 3:1. For the in-plane Fe-Nx active sites, there are usually defects in the surrounding carbon six-membered ring structure, resulting in an asymmetric structure of the Fe-N bond length, which is different from the traditionally recognized Fe-N4-C planar symmetric structure. The Fe-N4 bond length structure will have a significant impact on its catalytic properties and stability. The influence of Fe-N2 sites at the steps on the fuel cell oxygen reduction reaction was further investigated by density functional theory calculations. The atomic-scale coordination structure information analyzed in this paper will help to understand the catalytic active site structure and catalytic properties of Fe-N-C single-atom catalysts.