体块晶体中单个位错的自动伯格斯矢量识别
Automated Burgers Vector Identification for Individual Dislocations in Bulk Crystals
浏览论文内容
中文总结 AI 辅助
本文提出一种纳入晶体学约束的物理信息卷积神经网络，可从弱束积分摇摆曲线图像中自动识别体块晶体单个位错的伯格斯矢量，合成测试准确率达93%，实验交叉滑移案例中72.7%分层预测匹配参考伯格斯矢量。
中文摘要 AI 辅助
暗场X射线显微镜（DFXM）中的弱束成像可解析体块晶体中的单个位错，但从所得衬度分配伯格斯矢量通常需要与正演模拟进行手动对比。本文中，我们基于面心立方（FCC）铝中孤立位错的几何光学模拟，训练了一种物理信息卷积神经网络（CNN），将晶体学约束纳入学习过程。该模型可从弱束积分摇摆曲线图像中识别伯格斯矢量。在合成测试数据上，模型实现了约93%的准确率；在实验交叉滑移案例中，该受约束模型将72.7%的分层预测分配给参考伯格斯矢量。这些结果表明，经模拟训练的物理信息CNN是DFXM中自动位错识别的重要一步。
英文摘要
Weak-beam imaging in dark-field X-ray microscopy (DFXM) can resolve individual dislocations in bulk crystals, but assigning Burgers vectors from the resulting contrast typically requires manual comparison with forward simulations. Here, we train a physics-informed convolutional neural network (CNN) on geometrical optics simulations of isolated dislocations in face-centred cubic (FCC) aluminium, incorporating crystallographic constraints into the learning pro- cess. The model identifies Burgers vectors from weak-beam integrated rocking-curve images. On synthetic test data, the model achieves an accuracy of approximately 93%. In an experimental cross-slip case, the constrained model as- signs 72.7% of the layer-wise predictions to the reference Burgers vector. These results show that simulation-trained, physics-informed CNNs represent a step toward automated dislocation identification in DFXM.
