Machine learning-based beta transus temperature prediction for titanium alloys

Abstract

Beta transus temperature (βtr) is one of the most crucial features of titanium alloys. It is typically used as the index while designing the heat treatment process for titanium alloys. The βtr is also a significant parameter to optimize the processing technology of titanium alloys. Four machine learning algorithms and one empirical formula is developed in this study to estimate the βtr of titanium alloys: Artificial Neural Networks (ANN), Gauss Processing Regression (GPR), Super Vector Machine (SVM), and Ensemble Regression Trees (ERT). According to the correlation coefficient (R), Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) to verify the accuracy of models, and the experimentally measured phase transition temperature of Ti600 alloy was also used to verify the generalization ability of the model. Choosing the best model to analyze the sensitivity of the elements and determine how each component affects the βtr. The result demonstrated that the ANN model has the highest prediction accuracy among the five models, and different model structures have different effects on predicting new data. The ANN model with 10 neurons has the highest prediction accuracy, while the ANN model with 8 neurons has the strongest generalization ability. The results of the sensitivity analysis proved that all the alloy compositions used as input parameters were valid parameters.

Keywords

• Beta transus temperature
• Machine learning
• Titanium alloys
• Empirical formula
• Sensitivity analysis