<p>To address the challenges of low predictive accuracy, data sensitivity, and model complexity inherent in the Arrhenius phenomenological model for predicting the flow stress of twinning-induced plasticity (TWIP) and hot-formed steel (HFS) composites, a Bayesian-optimized XGBoost (BO-XGBoost) model was developed and rigorously validated through comparative analysis. The results demonstrate that the predictive performance of the BO-XGBoost model was significantly improved compared to the Arrhenius model. Specifically, the root mean square error decreased from 16.3160 to 1.0554, corresponding to an accuracy improvement of approximately 93.5%. Using the predicted flow stress values from the BO-XGBoost model, hot processing maps for the tested steel were constructed, and the microstructures under various deformation conditions were characterized in detail. The results indicate that, at high temperatures or low strain rates, the flow curves primarily exhibit recrystallization behavior. In contrast, at higher strain rates or lower temperatures, the flow curves display characteristics of work hardening. Specifically, multiple peak flow curves were observed at a strain rate of 0.1&#xa0;s⁻¹ and deformation temperatures ≤ 1050&#xa0;°C. The changes in the flow curves are attributed to the competition between work hardening, controlled by dislocation accumulation and interaction, and softening, governed by dynamic recovery and recrystallization. Furthermore, the hot processing maps reveal that the tested steel demonstrates optimal machinability within the deformation temperature range of 1075–1150&#xa0;°C and strain rate range of 0.05–0.5&#xa0;s⁻¹. This finding provides valuable insights for optimizing processing conditions and enhancing the material performance of TWIP and HFS composites in manufacturing and industrial applications.</p> Graphical Abstract