Prediction of mechanical properties and process optimization for Ti-Al-V-Fe-Mo-Cr titanium alloys using machine learning

Abstract

Titanium alloys have found extensive application in the aerospace field due to their outstanding properties. However, traditional material design methods involve extensive computational and experimental trial-and-error processes, leading to increased design costs and reduced efficiency. This paper proposes a machine learning-based design method for high-performance titanium alloys. It systematically evaluates the predictive capabilities of four “alloy composition-mechanical property” machine learning models, with the MLP model demonstrating optimal performance. The R2 values for predicting UTS and YS are 0.9516 and 0.8851, respectively. Feature engineering analysis revealed that increasing Mo, Al, and Cr contents enhances alloy strength. Combined with a multi-objective optimization Pareto front map, the Ti-2.5Al-2.2V-1.6Mo-1.8Fe-0.8Cr alloy was successfully identified as having optimal comprehensive mechanical properties. Further investigation revealed that annealing temperature and duration significantly influence the mechanical properties of the titanium alloy. As annealing temperature increases, strength initially decreases before rising, while elongation exhibits the opposite trend. Experimental validation ultimately confirmed that the annealing process at 750 °C for 30 minutes optimally balances strength and ductility in the target composition alloy, achieving an ultimate tensile strength (UTS) of 1176 MPa, yield strength (YS) of 1115 MPa, and elongation (EL) of 11.9%. This approach provides a novel pathway for the rapid design of high-performance titanium alloys. To improve physical interpretability, the revised feature framework additionally defines the Mo equivalent (Moeq), Al equivalent (Aleq), and the d-electron parameters Md and Bo; for the optimized nominal composition, Moeq = 9.54 wt. %, substitutional Aleq = 2.50 wt. %, Md = 2.391 eV, and Bo = 2.774. Quantitative image statistics from the available optical micrographs further give an exploratory Hall–Petch-type relation between YS and apparent α-domain size (R2 = 0.79). The reported machine-learning R2 values are explicitly retained as single hold-out estimates because the row-level 65-record dataset required to reconstruct numerical cross-validation statistics is not embedded in the manuscript file.