Physics-informed hybrid learning for dynamic constitutive and equation-of-state parameter identification with case-level ballistic assessment of titanium alloys

Reliable simulation of titanium-alloy impact response requires separate descriptions of deviatoric strength and volumetric shock compression. Here, a hybrid workflow identifies these two parameter groups from heterogeneous mechanical data and limited shock-compression data. We obtain JC parameters through unified curve-level calibration using a 604-curve benchmark from 14 alloys and 81 alloy states. The mean state-level relative calibration error across the 81 reported alloy states is 0.099, while the separate seven-curve TA15 state 1-0 calibration gives a pooled pointwise NMAE of 2.68% and a mean absolute stress deviation of 34.73 MPa. A curve-level held-out audit of the 604-curve benchmark gives a median NMAE of 9.60% (interquartile range 4.83-20.07%), and the separate TA15 state 1-0 dataset gives a mean held-out NMAE of 3.88%. For EOS reconstruction, the original decoder adaptation on 11 measured alloys remains a fitted-set analysis, with average differences of 1.05% for C₀ and 0.35% for λ. A separate nested leave-one-alloy-out audit using a frozen 32-dimensional pretrained feature representation and a 66-coefficient ridge readout gives held-out MAEs of 0.0719 km/s (1.41%) for C₀ and 0.00443 (0.414%) for λ. We examined the archived TA15 material card in two AUTODYN impact cases at 760 and 809 m/s, yielding mean absolute differences of 0.16 mm in penetration depth and 0.025 mm in back-face bulge. The results therefore distinguish in-sample calibration, held-out assessment, and case-level simulation, and show that performance depends on conditions and alloys rather than being universally predictive.