Adaptive characterization of microstructure dataset using a two stage machine learning approach
Adaptive characterization of microstructure dataset using a two stage machine learning approach
复制标题
DOI：
10.1016/j.commatsci.2020.109593
复制
发表时间：
2020-05-01
影响因子：
3.3
通讯作者：
Lewis, Daniel
中科院分区：
文献类型：
--
作者：
Baskaran, Arun;Kane, Genevieve;Lewis, Daniel
材料信息学（一般）和图像驱动机器学习（具体）的目标之一是从显微照片中提取定量数据，以有效表征微观结构特征。为了实现这一目标，我们报告了一个新的范例系统分割的形态特征相关的一个给定的微观结构，使用工业相关的钛合金作为一个例子。由分类步骤和分割步骤组成的两阶段流水线用于处理包含多个形态特征的钛微结构，并输出与所识别的微结构的特定类别相关的定量测量。对于分类步骤，使用Keras API训练卷积神经网络，其架构由三个卷积层和一个全连接层组成。显微结构分为三个标签：“层状”，“双相”，和“针状”。建立了1225幅图像的材料显微组织数据集，包括从七种不同的热处理条件下获得的Ti-6Al-4V合金显微照片。CNN在1000张图像的数据集上进行了训练，随后在225张图像的数据集上进行了测试。它报告了93.00 +/-1.17%的准确度，平均超过5次试验，将总数据集随机分为训练集和测试集。对于流水线的第二阶段，选择特定于显微照片的分类标签的图像处理技术。等轴晶粒的面积分数提取的双模态显微组织使用基于标记的分水岭技术，和面积分数的主要α-变量提取的篮子编织结构使用直方图的定向折射率（HOG）的方法。工程师可以使用类似于本工作中展示的概念验证管道的计算工具来更好地识别由于工艺或材料变化而产生的微结构特征。
One of the goals of Materials Informatics (generally) and Image Driven Machine Learning (specifically) is extraction of quantitative data from micrographs towards an efficient characterization of microstructural features. Towards this goal, we report on a new paradigm for systematic segmentation of morphological features relevant to a given microstructure using an industrially relevant titanium alloy as an example. A two stage pipeline consisting of a classification step and a segmentation step is used to process titanium microstructures containing multiple morphological features and output quantitative measurements relevant to the particular class of microstructure identified. For the classification step, a Convolutional Neural Network is trained using the Keras API, with the architecture consisting of three convolutional layers and one fully connected layer. The microstructures are classified into three labels: "lamellar", "duplex", and "acicular". A material microstructure dataset of 1225 images is established, comprised of Ti-6Al-4V alloy micrographs acquired from seven different thermal processing conditions. The CNN was trained on a dataset of 1000 images and subsequently tested on a dataset of 225 images. It reported an accuracy of 93.00 +/- 1.17%, averaged over 5 trials incorporating a random division of the total dataset into training and test sets. For the second stage of the pipeline, image processing techniques were selected specific to the classification label of the micrograph. The area fraction of equiaxed grains is extracted from bi-modal microstructures using a marker-based watershed technique, and the area fraction of the dominant alpha-variant is extracted from basket-weave structures using a Histogram of Oriented Gradients (HOG) method. Computational tools similar to the proof of concept pipeline demonstrated in this work can be used by engineers to better identify microstructural features that arise due to process or material variations.