Interpretable machine learning model for hot deformation behavior of Ti-Al-V-Fe-Cr armor titanium alloy

Hot deformation test of Ti-Al-V-Fe-Cr armor titanium alloy was conducted at 750–1025 °C and 0.01–10 s⁻¹. Using the friction and adiabatic corrected stress–strain curve data, modified Johnson-Cook (M-JC), genetic algorithm JC (GA-M-JC), and six Bayesian-optimized machine learning models were developed. Compared with M-JC, GA-M-JC significantly improved the fitting accuracy of strain-stress curve through genetic algorithm optimization. Taylor diagrams demonstrated that the GBR model achieved optimal correlation, outperforming LSTM, RF, 1D-CNN, BPNN and SVR. Compared with M-JC and GA-M-JC, the GBR had the lowest extrapolation errors under hot deformation extrapolation conditions. SHAP analysis showed that strain rate was the dominant flow-stress factor. The high-stress and low-stress regions predicted by GBR was aligned with the microstructural changes. It indicated that the model followed the physically deformation mechanisms beyond the experimental domain and offered an interpretable basis for hot deformation prediction of Ti-Al-V-Fe-Cr alloy.