Machine learning in designing amorphous alloys

Metallic glasses have received a lot of interest because of their excellent mechanical, physical, and chemical qualities. For example, they have a stronger resistivity than crystalline metals composed of the same elements and a lower viscosity coefficient. However, the difficulty in creating alloy compositions has been a concern for researchers. Traditional amorphous alloy systems design approaches, such as empirical trial-and-error methods and methods based on density functional theory (DFT), have assisted researchers in exploring numerous amorphous alloy systems during the growth of materials science over the last few decades. However, with the continuous development of materials science, these methods have been difficult to meet the needs of researchers due to their long development cycles and low efficiency. Additionally, the complex and long-range disordered structure of metallic glasses makes it difficult to understand their structure and nature in a comprehensive and clear way by conventional methods. Amorphous alloy composition design and property analysis are now often conducted using machine learning techniques because of their low experimental cost, short development cycle, strong data processing capability, and high predictive performance, among other advantages. They present new approaches and chances to address significant key bottlenecks in the field of metallic glass. In this study, the main processes of machine learning model building were introduced. Subsequently, the related studies on data pre-processing, model construction, and model validation were presented. For data pre-processing, data selection, feature engineering, and advanced data balancing methods were primarily described. In the feature engineering part, the model performance with various input features was examined, and it was shown that either employing physical properties or directly using the alloy compositions as the model input might result in high performance. Four machine learning algorithms were used to generate the machine learning model: artificial neural networks (ANN), support vector machines (SVM), random forest (RF), and extreme gradient boosting (XGBoost). A comparison indicates that SVM models work best with small data sets, whereas the performance of all other models tends to get better as the amount of training data increases. Generally, the XGBoost method outperforms several other methods and is, therefore, often used in machine learning competitions. Model validation approaches: K-fold cross-validation and leave-one-out cross-validation methods were presented. A good metallic glass performance prediction method needs to perform well in both validation methods. Finally, this study provides several possible future research directions on feature engineering, dataset construction, validation, and machine learning models.