Machine learning prediction of magnetic properties of Fe-based metallic glasses considering glass forming ability. (20th March 2022)

• Record Type: Journal Article
• Authors: Li, Xin; Shan, Guangcun; Shek, C.H.
• Is Part Of: Journal of materials science & technology. Volume 103(2022)
• Page Start: 113
• Page End: 120
• Publication Date: 2022-03-20
• Subjects: Metallic glasses -- Soft magnetic properties -- Glass forming ability -- Machine learning -- Non-linear regression; Metals -- Periodicals; Materials science -- Periodicals; Materials science; Metals; Periodicals; 620.1105
• DOI: 10.1016/j.jmst.2021.05.076
• Languages: English
• ISSNs: 1005-0302

Abstract

Abstract: Fe-based metallic glasses (MGs) have shown great commercial values due to their excellent soft magnetic properties. Magnetism prediction with consideration of glass forming ability (GFA) is of great significance for developing novel functional Fe-based MGs. However, theories or models established based on condensed matter physics exhibit limited accuracy and some exceptions. In this work, based on 618 Fe-based MGs samples collected from published works, machine learning (ML) models were well trained to predict saturated magnetization ( B s ) of Fe-based MGs. GFA was treated as a feature using the experimental data of the supercooled liquid region (Δ T x ). Three ML algorithms, namely eXtreme gradient boosting (XGBoost), artificial neural networks (ANN) and random forest (RF), were studied. Through feature selection and hyperparameter tuning, XGBoost showed the best predictive performance on the randomly split test dataset with determination coefficient ( R 2 ) of 0.942, mean absolute percent error (MAPE) of 5.563%, and root mean squared error (RMSE) of 0.078 T. A variety of feature importance rankings derived by XGBoost models showed that Δ T x played an important role in the predictive performance of the models. This work showed the proposed ML method can simultaneously aggregate GFA and other features in thermodynamics, kinetics and structures to predict the magnetic properties of Fe-based MGs with excellent accuracy. Highlights: A data-driven machine learning method was proposed to predict magnetic properties of Fe-based metallic glasses with consideration of GFA. Through feature selection and hyperparameter tuning, XGBoost showed the best predictive performance on the randomly split test dataset with R 2 of 0.942, MAPE of 5.563%, and RMSE of 0.078 T. A variety of feature importance rankings derived by XGBoost models showed that Δ T x played an important role in the predictive performance of the models, which indicated the correlation between GFA and B s .