Predicting elastic strain fields in defective microstructures using image colorization algorithms
Predicting elastic strain fields in defective microstructures using image colorization algorithms
复制标题
DOI：
10.1016/j.commatsci.2020.110068
复制
发表时间：
2021-01-01
影响因子：
3.3
通讯作者：
Basu, Saurabh
中科院分区：
文献类型：
--
作者：
Khanolkar, Pranav Milind;McComb, Christopher Carson;Basu, Saurabh
在这项工作中，基于卷积神经网络的图像彩色化算法，探索作为一种方法来预测拉伸平面应变场组件的微结构具有孔隙缺陷。同样，在ASTM-E8尺寸的数值试样的计量截面上对具有各种形状、尺寸、面积分数和数密度的孔隙的微观结构进行取样，使用商业有限元分析软件Abaqus在平面应变模式下模拟其拉伸变形。随后，将具有孔隙缺陷的显微组织作为灰度图像，将其应变场分量作为其颜色层，类似于传统数字图像的红-绿-蓝颜色分量，对图像彩色化算法进行训练。同样，测试了各种CNN框架以优化其参数，即层数，每层中的过滤器数量，步幅，填充和激活函数。提出了一种优化的CNN框架，该框架能够在有限元分析所需时间的一小部分内以高精度R-2 > 0.91预测随机采样微结构上的应变场。进行了各种交叉验证测试，以测试CNN在学习各种微结构特征时的准确性和鲁棒性。结果表明，CNN算法是非常强大的，可以在一般情况下提供接近准确的应变场。
In this work, an image colorization algorithm based on convolutional neural networks is explored as an approach to predict tensile plane-strain field components of microstructures featuring porosity defects. For the same, microstructures featuring porosity of various shapes, sizes, area fractions and number densities were sampled on the gage section of ASTM-E8 sized numerical specimens whose tensile deformation was simulated in plane strain mode using commercial finite element analysis package Abaqus. Subsequently, the image colorization algorithm was trained by treating the microstructure featuring porosity defects as the gray scale image, and its strain field components as its color layers, analogous to the red-green-blue color components of traditional digital representations of images. Towards the same, various CNN frameworks were tested for optimization of its parameters, viz. number of layers, number of filters in each layer, stride, padding, and activation function. An optimized CNN framework is presented that is able to predict strain fields on randomly sampled microstructures with high accuracy R-2 > 0.91 at a fraction of the time that finite element analysis would take. Various cross-validation tests were performed to test the accuracy and robustness of the CNN in learning features of various microstructures. Results indicated that the CNN algorithm is extremely robust and can provide near-accurate strain fields in generic scenarios.