<p>High-temperature compression is an effective approach for regulating the microstructure and achieving grain refinement of titanium alloys. In this study, the constitutive relationship of Ti65 alloy was established using two complementary models: the strain-compensated Arrhenius (SCA) model and a backpropagation artificial neural network (BPANN). The accuracy and reliability of the developed models were systematically evaluated via statistical analysis and cross-validation. Results indicated that the BPANN model outperformed the SCA model, demonstrating higher prediction accuracy and superior generalization capability for describing the hot deformation behavior of Ti65 alloy. Subsequently, the finite element (FE) software Deform-3D was employed to simulate the hot compression process of Ti65 alloy, with a specific focus on investigating grain size evolution and dynamic recrystallization (DRX) behavior. The effects of key deformation parameters—including reduction ratio, deformation speed, and initial deformation temperature—on grain refinement and DRX behavior were quantitatively revealed. Through comprehensive analysis of average grain size, DRX volume fraction, and equivalent strain distribution, the optimal deformation parameters for Ti65 alloy were determined as a deformation speed of 5&#xa0;mm/s and an initial temperature of 1020&#xa0;°C. This work provides valuable theoretical guidance and technical support for the optimization of hot deformation parameters of Ti65 alloy, thereby facilitating the precise control of its microstructure and mechanical properties in industrial applications.</p> Graphical abstract <p></p>