Composition and structure analyzer/featurizer for explainable machine-learning models to predict solid state structures

Abstract

Traditional and non-classical machine learning models for solid-state structure prediction have predominantly relied on compositional features (derived from properties of constituent elements) to predict the existence of a structure and its properties. However, the lack of structural information can be a source of suboptimal property mapping and increased predictive uncertainty. To address this challenge, we have introduced a strategy that generates and combines both compositional and structural features with minimal programming expertise required. Our approach utilizes open-source, interactive Python programs named Composition Analyzer Featurizer (CAF) and Structure Analyzer Featurizer (SAF). CAF generates numerical compositional features from a list of formulae provided in an Excel file, while SAF extracts numerical structural features from a .cif file by generating a supercell. 133 features from CAF and 94 features from SAF are used either individually or in combination to cluster nine structure types in equiatomic AB intermetallics. The performance is comparable to those with features from JARVIS, MAGPIE, mat2vec, and OLED datasets in PLS-DA, SVM, and XGBoost models. Our SAF + CAF features provide a cost-efficient and reliable solution, even with the PLS-DA method, where a significant fraction of the most contributing features is the same as those identified in the more computationally intensive XGBoost models.