An ensemble learning based amorphous state predictor for multicomponent alloys

Metallic glasses have received extensive academic attention due to their unique characteristics, such as soft magnetic properties, superconductivity, corrosion resistance, high strength, and hardness. However, its limited glass-forming ability hampers the design of new metallic glass systems. The machine learning models can predict the properties and phase state of an alloy based on the existing experimental data, helping researchers design new alloy systems. Model selection and data balancing are crucial for machine learning models to prevent overfitting and improve model performance. In this paper, we used an ensemble learning model with the artificial neural network and random forest as its base models to obtain better results. Moreover, we proposed a data balancing method to improve the performance of our model for different data distributions by balancing the data distribution between alloy systems and within each alloy system. Finally, compared with other machine learning models, we achieve at least 1.2% and 2.2% loss reduction for known alloy systems with a new composition ratio and new alloy systems, separately. And our data balancing method has a 3.6% performance improvement over the state-of-the-art data balancing method.