Integrating machine learning and CALPHAD method for exploring low‐modulus near‐β‐Ti alloys
Hao Zou, Yueyan Tian, Ligang Zhang, Renhao Xue, Zixuan Deng, Mingming Lu, Jianxin Wang, Libin Liu
Abstract
Abstract Traditional theoretical and empirical calculation methods can guide the design of β‐ and metastable β‐alloys for bio‐titanium. However, it is still difficult to obtain novel near‐β‐Ti alloys with low modulus. This study developed a method that combines machine learning with calculation of phase diagrams (CALPHAD) to facilitate the design of near‐β‐Ti alloys. An elastic modulus database of Ti–Nb–Zr–Mo–Ta–Sn system was constructed first, and then three features (the electron to atom ratio, mean absolute deviation of atom mass, and mean electronegativity) were selected as the key factors of modulus by performing a three‐step feature selection. With these features, a highly accurate model was built for predicting the modulus of near‐β‐Ti alloys. To further ensure the accuracy of modulus prediction, machine learning with the elastic constants calculated was leveraged by CALPHAD database. The root mean square error of the well‐trained model can be as low as 6.75 GPa. Guided by the prediction of machine learning and CALPHAD, three novel near‐β‐Ti alloys with elastic modulus below 50 GPa were successfully designed in this study. The best candidate alloy (Ti–26Nb–4Zr–4Sn–1Mo–Ta) exhibits an ultra‐low modulus (36.6 GPa) after cold rolling with a thickness reduction of 20%. Our method can greatly save time and resources in the development of novel Ti alloys, and experimental verifications have demonstrated the reliability of this method.