A machine learning method approach for designing novel high strength and plasticity metastable β titanium alloys

In order to improve efficiency and reduce costs, four machine learning models were established for the design of metastable β titanium alloys, including the Adaboost model, the LightGBM model, the Voting model, and the Stacking model. The accuracy of these models was evaluated, and all models exhibited excellent accuracy for tensile strength, yield strength, and elongation. The values of R-squared coefficients(R²) all greater than 0.9. Among them, the LightGBM model showed the highest accuracy, with relatively smallest values of mean absolute error (MAE) and root mean square error (RMSE) and relatively largest value of R². To further verify the accuracy of the model, a metastable β titanium alloy Ti-5.5Cr-5Al-4Mo-3Nb-2Zr was designed by the LightGBM model. The predicted values of the alloy's tensile strength, yield strength, and elongation under three heat treatment processes were in high agreement with the experimental values. The alloy exhibited optimal strength-plasticity matching after undergoing a solution treatment at 850 ​°C for 0.5 ​h, followed by aging at 650 ​°C for 8 ​h, with a tensile strength of 1317 ​MPa, an elongation of 11.17 ​%, and a strength-plasticity product of 14.711 ​GPa·%.