Optimizing the hot working processes of advanced titanium alloys is crucial for aerospace applications, yet predictive modeling remains a significant hurdle, especially for novel compositions. This study addresses this challenge by developing an integrated framework that combines machine learning with multi-scale microstructural analysis to predict and explain the hot deformation behavior of a TC4 alloy microalloyed with Cerium (Ce). A comparative evaluation of constitutive models showed that an XGBoost machine learning algorithm achieved superior predictive fidelity (R 2 = 97.5%, RMSE = 6.98), substantially surpassing the conventional phenomenological Arrhenius model (R 2 = 87.5%, RMSE = 18.38). The model’s predictive logic, as confirmed by SHAP interpretability analysis, aligns with fundamental metallurgical principles like strain rate hardening and thermal softening. The primary dynamic softening mechanism was identified as dynamic recrystallization (DRX) through Electron Backscatter Diffraction (EBSD), which operates via both discontinuous (DDRX) and continuous (CDRX) pathways. Furthermore, high-magnification Transmission Electron Microscopy (TEM) revealed that the enhanced performance from Ce microalloying is attributed to dispersed CeO 2 nanoparticles at α/β phase interfaces. These nanoparticles impede dislocation movement, creating localized high-strain-energy zones that act as preferential sites for DRX grain nucleation, thereby accelerating DRX kinetics and refining the grain structure. This work not only establishes a high-fidelity predictive tool for titanium alloy processing but also provides an interpretable link between the model and its underlying physical mechanisms, validated by experimental results. The approach highlights a powerful synergy between materials science and machine learning, paving the way for the intelligent design of next-generation alloys.
Li et al. (2026) studied this question.