清华给电子显微镜加上Agent，DeepSeek V3全程调度，数天流程缩短至几分钟

Core Viewpoint - The article discusses the launch of AutoMat, an AI agent developed by Tsinghua University and other institutions, which automates the process of converting atomic-level STEM images into standard CIF structures, significantly reducing the time required for material analysis and discovery [1][2][5]. Group 1: AI Agent Functionality - AutoMat acts as a precise "map translator," transforming atomic-level STEM images into standard CIF structures and providing key physical properties like formation energy in one step [2]. - The traditional manual process of analyzing images, which could take hours or days, has been reduced to a few minutes, effectively bridging the gap between "microscopic imaging" and "structural reconstruction" [3][5]. - The team created a dataset called STEM2Mat-Bench, consisting of over 450 samples of two-dimensional materials, to validate AutoMat's performance [3][7]. Group 2: Methodology and Performance - AutoMat operates in four main steps: noise reduction, structure prior retrieval, atomic reconstruction, and property prediction, allowing for a closed-loop process from image to material insight [11][16]. - The performance evaluation of AutoMat against existing models shows it significantly outperforms multimodal models and specialized tools like AtomAI in both reconstruction accuracy and energy prediction [19][21]. - AutoMat achieved an average absolute error of 332 ± 12 meV/atom in formation energy prediction, which, while higher than the theoretical best of 48 meV, is substantially lower than the errors typically seen in visual-language models [20][21]. Group 3: Challenges and Future Directions - The article identifies two main challenges: template retrieval failures (39.3%) and downstream reconstruction failures (60.7%), which can lead to significant errors in atomic arrangement and structure output [22]. - Future improvements will focus on integrating experimental, simulation, and AI processes to enhance the accuracy of complex material structure and property predictions [23].