Data Science Techniques, Assumptions, and Challenges in Alloy Clustering and Property Prediction

Abstract

Data analytics methods have been increasingly applied to understanding materials chemistry, processing due to the manufacturing approach, and uni-axial and cyclic property relationships in the highly complex space of alloy design. There are several benefits to applying data analytics to this space, including the ability to manage non-linearities in the responses of the alloy attributes and the resulting mechanical properties. However, key difficulties in applying and understanding the results of data analytics include the often lack of reported assumptions and data processing steps necessary to improve interpretation and reproducibility in derived results. In this work, the methods used to generate clustering and correlation analyses for experimental 9% Cr ferritic-martensitic steel data were investigated and the resulting implications for mechanical property predictions were assessed. This work uses principal component analysis, partitioning around medoids, t-SNE, and k-means clustering to investigate trends in composition, processing and microstructure information with creep and tensile properties, building on work done previously using a smaller version of the same dataset. The initial assumptions, preprocessing steps and methods are investigated and outlined in order to depict the fine level of detail required to convey the steps taken to process data and produce analytical results. The variations in the resulting analyses are explored due to the influence of new and more varied data.

Assessment of Outliers in Alloy Datasets Using Unsupervised Techniques

“…This approach was based on previous work using a similar dataset. [58][59][60] Using this domain knowledge-based approach to estimate the number of clusters improves the interpretability of the clustering model, as well as the outlier model, by adding physics-and materials-based insights into the clustering patterns. Further discussion of the techniques used to choose the number of clusters is included in the supplementary material.…”

“…1a, with color applied based on the k-means cluster to which each point belongs. Using the labeling technique described in previous works, [58][59][60] 14 clusters composed the 9-12% Cr dataset based on alloy composition and heat treatment values. As this method was based on domain knowledge of the underlying attributes of the alloy class, this method was used to determine the number of clusters needed to visualize and understand outliers in the data.…”