Phase prediction and experimental realisation of a new high entropy alloy using machine learning

Area of Science:

• Materials Science and Engineering
• Computational Materials Science
• Machine Learning Applications

Background:

• High entropy alloys (HEAs) offer vast compositional possibilities (~10^8 types) but predicting their structure and properties is challenging.
• Experimental methods for HEA development are time-consuming and resource-intensive, necessitating predictive tools.
• Existing literature often uses disparate synthesis routes, leading to data discrepancies and erroneous conclusions in machine learning models.

Purpose of the Study:

• To investigate the efficacy of machine learning (ML) algorithms for predicting the crystal structure of high entropy alloys (HEAs).
• To evaluate the reliability of different ML models, including K-nearest neighbours (KNN), support vector machine (SVM), decision tree classifier (DTC), random forest classifier (RFC), and XGBoost (XGB).
• To assess the impact of synthetic data augmentation techniques on the predictive accuracy of ML models in HEA development.

Main Methods:

• Screened a large dataset of experimentally fabricated HEAs using melting and casting methods to ensure data consistency.
• Tested five base machine learning algorithms (KNN, SVM, DTC, RFC, XGB) on the curated dataset.
• Compared the performance of a vanilla Random Forest Classifier (V-RFC) on original data against a synthetic minority over-sampling technique-Tomek links (SMOTE-Tomek) augmented RFC (ST-RFC) model.

Main Results:

• The Random Forest Classifier (RFC) demonstrated superior prediction reliability compared to other tested algorithms.
• Synthetic data augmentation (SMOTE-Tomek) did not yield significant improvements in phase prediction accuracy for HEAs, despite a higher average test accuracy (92%).
• Analysis using confusion matrices and ROC-AUC scores indicated that augmented data did not lead to breakthroughs in predicting individual phases.

Conclusions:

• Machine learning, particularly the RFC model, is a viable and robust tool for predicting HEA crystal structures and facilitating the discovery of novel alloys.
• Synthetic data augmentation is not recommended for HEA development due to its unreliability in assuring accurate phase information.
• The study successfully predicted a new HEA (Ni25Cu18.75Fe25Co25Al6.25) with a face-centered cubic (FCC) phase, validating the RFC model's predictive power.