Physical-information machine learning for strength and ductility prediction of metastable β titanium alloys

Machine learning for predicting metastable β titanium alloy properties suffers from limited generalisation, unreliable extrapolation beyond data boundaries, and inadequate modelling of nonlinear effects of process parameters. This study proposes a physics-informed machine learning approach incorporating intrinsic physical attributes and phase transformation kinetics. Using 496 samples, it achieves R2 = 0.95 for ultimate tensile strength (UTS) and 0.90 for elongation (El), accurately predicting out-of-boundary alloys with similar heat treatment responses within 5.0% UTS and 2.5% El error. This approach reduces reliance on data completeness and complex algorithms, providing an accurate, generalisable path to UTS- and El-targeted titanium alloy design.