Abstract The application of machine learning and computer vision approaches to microscale deformation data for the automated identification of deformation twinning systems, and their associated twin area, is introduced. Deformation data was obtained during in-situ SEM compression testing of a WE43 Mg alloy using digital image correlation (DIC) modified for use with electron microscopy. A 5.7 mm × 3.4 mm area of interest was analyzed, generating ~100 million data points of deformation. Experimental twin trace directions were determined by applying k-means clustering, morphological thinning, and Hough transforms to the deformation data. The identification of deformation twinning was achieved by consideration of the twin trace direction, its strain value, and its evolution. The deformation map was divided into areas corresponding to individual active twin systems, enabling the analysis of microstructure dependence of deformation behavior and twinning activity. The performance of the proposed twinning identification approach was evaluated by accuracy, precision, and sensitivity metrics. The effect of the tuning parameters used in the algorithm on the performance is also discussed.

中文翻译：

---

从 SEM DIC 自动识别镁 WE43 中的变形孪生系统

摘要介绍了机器学习和计算机视觉方法在微尺度变形数据中的应用，用于自动识别变形孪生系统及其相关的孪生区域。变形数据是在 WE43 镁合金的原位 SEM 压缩测试过程中获得的，该测试使用经过修改以用于电子显微镜的数字图像相关 (DIC)。分析了 5.7 mm × 3.4 mm 的感兴趣区域，生成了约 1 亿个变形数据点。通过对变形数据应用 k 均值聚类、形态细化和霍夫变换来确定实验双轨迹方向。变形孪生的识别是通过考虑孪生轨迹方向、其应变值及其演变来实现的。变形图被划分为对应于单个活动孪生系统的区域，从而能够分析变形行为和孪生活动的微观结构依赖性。所提出的孪生识别方法的性能通过准确性、精度和灵敏度指标进行评估。还讨论了算法中使用的调整参数对性能的影响。