Research On Glass Forming Ability Of Amorphous Alloy Based On Machine Learning

Amorphous alloys are widely used in many fields due to their excellent physical,chemical,and mechanical properties and have received extensive attention.However,it is difficult to predict the glass-forming ability（GFA）of amorphous alloys efficiently and accurately,which limits the industrialization of amorphous alloys.Since there is currently no relevant theory to describe the GFA of amorphous alloys,only a summary based on limited experimental results,it is of great theoretical and practical significance to predict the GFA of amorphous alloys by using machine learning methods.This paper studies the GFA prediction of binary and multi-component alloys.In the study for the GFA of binary alloys,a Random Forest Regression model was used to fill in the absent but necessary feature in the predicting dataset to build a complete datasets.Using datasets with input features of different dimensions to evaluate six machine learning models,it was found that all models performed best with 9-dimensional input features,and the Gradient boosting decision tree（GBDT）model with the best predictive performance was used to predict GFA of binary alloys.The analysis shows that the feature has the greatest influence on the GFA of binary alloys is the liquidus temperature of the binary alloy(T_liq),followed by the solvent atomic fraction（c₁）,liquidus temperature difference(ΔT_liq),and atomic weight difference（W）.About 72.48%of binary alloys with good GFAs in the prediction dataset satisfy the empirical criterion of the atomic size difference R≥0.12.By studying the attribute independence assumption of Na（?）ve bayes（NB）algorithm,it is found that normalizing the data may affect the effectiveness of the NB algorithm.The GFA of multi-component amorphous alloys was studied with the maximum critical diameter(D_max)as the prediction target.We collected 1404 alloy samples from published papers and used their compositional information,characteristic temperature,and geometric features as input features.Three machine learning algorithms were used to fit the dataset.It was found that the XGBoost model can get the best prediction effect,and it is 11.20%higher than the best performance in published papers,indicating that the input features used in this experiment can effectively represent the GFA of multi-component alloys.To further optimize the input features of multi-component alloys,seven datasets with different input features were established.Through the comparison of multiple groups of ablation experiments,it was found that when the composition information and characteristic temperature were used as input features,the R²prediction performance of the XGBoost model could be further improved 0.7332,while indicating that the geometric features are hidden in the composition information.