Identify Survived Key Features and Relevant Mechanisms for Designing High-Entropy Carbides via AI or Machine Learning

Multielement high-entropy carbides
(HECs) provide many
opportunities
for HECs to obtain optimal combinations of various properties, e.g.,
high strength and high flexibility, leading to high toughness. However,
the multielements significantly increase the compositional arrangements,
challenging the development of advanced HECs. Machine learning (ML)
provides a powerful approach to HEC design/discovery. Identifying
key parameters or selecting the key features is crucial for carbide
design with desirable properties. In the meantime, developing a reliable
ML model with minimized mutual interference from multiple features,
toward more accurate property predictions, also benefits the carbide
discovery. In this study, we use a small carbide database to study
the correlations between elastic moduli and 13 features with the assistance
of recursive feature elimination (RFE) and investigate how the feature
selection affects the prediction of HECs’ properties. It is
demonstrated that the mutual interference among highly correlated
features may have a negative influence on the accuracy of ML prediction
due to their mutual interference or redundant noise. For HECs, a few
basic features are identified, which largely determine their elastic
moduli. Among them, electron work function and valence electron concentration
(VEC) appear to be more responsible for bulk and shear/Young’s
moduli, respectively. Other parameters are crucial for all elastic
moduli, such as mixing entropy, formation energy, bond order, and
bond length. This study demonstrates the significance of identifying
the prominent features with lowered mutual interference or noise in
further HECs with optimal properties.