Application of a Chained-ANN for Learning the Process–Structure Mapping in Mg2SixSn1−x Spinodal Decomposition
Application of a Chained-ANN for Learning the Process–Structure Mapping in Mg2SixSn1−x Spinodal Decomposition
复制标题
DOI：
10.1007/s40192-022-00274-3
复制
发表时间：
2022-09
影响因子：
3.3
通讯作者：
Grayson H. Harrington;Conlain Kelly;V. Attari;R. Arróyave;S. Kalidindi
中科院分区：
文献类型：
--
作者：
Grayson H. Harrington;Conlain Kelly;V. Attari;R. Arróyave;S. Kalidindi
该工作建立了可靠和准确的材料工艺结构(PS)代理模型，该模型将18维工艺参数输入域映射到单相和双相微结构的高维空间。这是通过使用材料知识系统(MKS)框架(包括通过两点统计进行微结构量化和使用主成分分析进行降维)来实现的，并随后构建了一个链式人工神经网络(ANN)来学习高维输入区域与相应微结构空间(包括均匀和非均匀微结构)的MKS派生低维表示之间的复杂非线性映射。这一工作流程的好处在 ~ 10,000最终微结构的集合上得到，这些微结构是从镁2SixSn1-x材料系统中的化学机械调幅分解相场模拟获得的。具体地说，所选案例研究的复杂相场过程结构关系可以用一个只有742个可拟合参数的稳健模型来捕捉。
This work establishes a reliable and accurate materials process–structure (PS) surrogate model that maps an 18-dimensional process parameter input domain to a high-dimensional space of single- and dual-phase microstructures. This was accomplished by employing the Materials Knowledge Systems (MKS) framework (includes microstructure quantification via two-point statistics and dimensionality reduction using principal components analysis) for the feature engineering of the microstructures, and subsequently constructing a chained-artificial neural network (ANN) to learn the complex nonlinear mappings between the high-dimensional input domain and the MKS-derived low-dimensional representation of the corresponding microstructure space (includes both homogeneous and heterogeneous microstructures). The benefits of this workflow are demonstrated on a collection of ~ 10,000 final microstructures obtained from chemo-mechanical spinodal decomposition phase-field simulations in the Mg2SixSn1-xmaterial system. Specifically, it is shown that the complex phase-field process–structure relationships for the selected case study can be captured in a robust model with only 742 fittable parameters.