Advancing constitutive analysis of the hot deformation behavior of Ti–3Al–8V–6Cr–4Mo–4Zr via physics-aided machine learning

Near- β Ti-3Al-8V-6Cr-4Mo-4Zr alloy necessitates an optimized thermomechanical process to ensure excellent hot workability. The traditional constitutive approach requires a large amount of hot deformation data and has the issue of insubstantially fixed parameters that limit its predictive performance and range of applications. On the other hand, a purely data-driven approach utilizing machine learning (ML) lacks generalizability and physical interpretability, particularly under extrapolated conditions. To address these limitations, this study proposes a physics-aided ML framework that combines the Zener–Hollomon equation with a hyperparameter-tuned neural network to dynamically determine material-dependent parameters with respect to the deformation temperature, strain rate, and plastic strain. The proposed framework successfully reduces the predictive error by 44.2% compared with the conventional Zener–Hollomon constitutive model in the most difficult predictive scenarios, including the lowest temperature and highest strain rate. Moreover, the proposed framework improves not only the predictive accuracy but also the generalizability, both for interpolation and extrapolation, while retaining physical transparency. To validate the predictive applicability, the proposed framework is employed to correct an incomplete processing map that reasonably explains the microstructural evolution under the given processing conditions.