A Clustering Optimization Framework for Quantification of Non-Uniform Microstructures Based on Two-Point Statistics and Principal Component Analysis

Abstract

Microstructure quantification is one of the core procedures in building process-structure-property linkages to accelerate the design and development of new materials. The approach of the two-point statistics technique combined with principal component analysis (TST-PCA) has emerged as an efficient and unbiased method for quantitatively capturing low-dimensional features of microstructures. However, when encountering a dataset with non-uniform microstructures, the traditional metric for evaluation can't precisely quantify the representativeness of low-dimensional features in terms of replacing the original microstructures, which makes it difficult in determining the proper number of reserved low-dimensional features. The information loss after data dimension reduction may be so great for partial members in the dataset that the low-dimensional features could not represent the original microstructures. To address this challenge, a new quantitative metric linked to the amount of structural information of microstructures captured by low-dimensional features was first designed. Next, a clustering optimization framework for the quantification of non-uniform microstructures based on TST-PCA was proposed. Based on the invariant moment method, a comparison between original and reconstructed microstructures was completed, which proves that the new metric is reasonable to measure the reliability of the improved microstructure quantification framework. Finally, The representative low-dimensional features of experimental microstructures were successfully obtained when the new framework was applied to an unbalanced dataset. This computationally convenient framework was envisioned to be applicable to most microstructure datasets.