STEMGym
收藏资源简介:
STEMGym数据集是为STEMGym基准测试提供的配套数据,用于评估自主剂量高效的扫描透射电子显微镜(STEM)代理。数据集包含模拟的STEM样本,以HDF5世界文件形式存储。每个世界文件包含:低倍全样本概览图像、高分辨率STEM图像(128×128像素的瓦片,4像素重叠,步长124,排列成8×8网格)、原子位置、缺陷类型和相图的真实标注,以及像素大小、加速电压、探测器几何形状和材料参数等元数据。数据集涵盖三种材料(SrTiO₃、BaTiO₃、SiGe),每种材料提供三个难度级别(简单、中等、困难),控制缺陷密度、噪声水平和空间分布。文件大小约为50MB每个,总规模在1K到10K之间。数据集适用于强化学习任务,特别是STEM显微镜的自主控制与优化。
The STEMGym dataset is supporting data for the STEMGym benchmark, designed to evaluate autonomous dose-efficient scanning transmission electron microscopy (STEM) agents. The dataset contains simulated STEM samples, stored in HDF5 world files. Each world file includes: low-magnification full-sample overview images, high-resolution STEM images (128×128 pixel tiles with 4-pixel overlap, step size 124, arranged into an 8×8 grid), ground-truth annotations of atomic positions, defect types and phase diagrams, as well as metadata such as pixel size, acceleration voltage, detector geometry and material parameters. The dataset covers three materials (SrTiO₃, BaTiO₃, SiGe), with three difficulty levels (easy, medium, hard) provided for each material, which control defect density, noise level and spatial distribution. Each file is approximately 50 MB in size, and the total scale ranges from 1K to 10K. This dataset is applicable to reinforcement learning tasks, especially autonomous control and optimization of STEM microscopes.



