A Fully Annotated Synthetic Dataset of Fiber Microstructures for Machine Learning–Based Orientation Estimation

This repository contains a large-scale synthetic dataset of fiber microstructures designed for training, validating, and benchmarking machine-learning models for orientation tensor estimation in short-fiber reinforced polymers. The dataset consists of tens of thousands of microscopy-like binary images generated using a physics-informed simulation pipeline that replicates the morphological characteristics of composites processed via Large-Format Additive Manufacturing (LFAM). Each synthetic image is paired with a CSV file containing per-fiber geometric descriptors and the analytically computed second-order orientation tensor for the entire microstructure. Because the underlying fiber geometry is known exactly, the dataset provides noise-free ground-truth annotations, enabling rigorous and reproducible machine-learning research. The dataset is intended to support applications in materials science, computer vision, composite characterization, digital twins, and data-driven manufacturing.本仓库包含一套大规模纤维微观结构合成数据集，该数据集专为短纤维增强聚合物（short-fiber reinforced polymers）体系中取向张量（orientation tensor）估计任务的机器学习模型训练、验证与基准测试而设计。该数据集包含数万幅类显微二值图像，均通过物理信息驱动的仿真流程生成，可复现采用大尺寸增材制造（Large-Format Additive Manufacturing, LFAM）工艺制备的复合材料的形貌特征。 每幅合成图像均配套一份CSV文件，其中包含单根纤维的几何描述符，以及针对整个微观结构的解析计算二阶取向张量。由于纤维的底层几何结构完全已知，该数据集提供无噪声的真实标注（ground-truth annotations），可支撑严谨且可复现的机器学习研究。 本数据集旨在支撑材料科学、计算机视觉、复合材料表征、数字孪生及数据驱动制造等领域的相关应用。