Abstract: The impact mechanical properties of metallic materials are intrinsically correlated with applied loading strain rate and service temperature, which significantly govern their deformation behavior under dynamic conditions. Owing to its concise mathematical formulation, the classical Johnson-Cook (J-C) constitutive model has been extensively and successfully employed to describe the impact response of various metals. However, the inherent decoupling assumption in the J-C model, which independently considers strain hardening, strain-rate hardening, and thermal softening, frequently yields considerable deviations and insufficient characterization accuracy when materials are subjected to coupled wide-range strain rates (from quasi-static to high-strain-rate regimes) and broad temperature spans. To overcome these limitations, this paper proposes a rate-temperature coupled impact constitutive characterization method that integrates J-C constitutive constraints with data-driven techniques. Systematic impact mechanical experiments were conducted on two representative alloys: 7075-T6 aluminum alloy and 65Mn high-carbon steel. The tests covered the temperature range from room temperature (RT) to 500 °C and the strain-rate range of 10⁻³ to 10³ s⁻¹. Based on the experimental data, a grey-box Physics-Informed Neural Network (PINN) constitutive model embedded with J-C model was established, and a comparative investigation was then performed among the PINN, the traditional J-C model, and a standard Artificial Neural Network (ANN) to evaluate their respective predictive capabilities in stress prediction for metallic materials. The results show that the PINN reconstructs input features by decomposing strain contributions into the strain-rate hardening term, the temperature softening term and rate-temperature coupling term for neural-network training. Compared with J-C and ANN models, the PINN achieves the best overall prediction accuracy. The maximum relative Root-Mean-Square Error (rRMSE) of the PINN model for all loading conditions of the two metallic materials within the temperature range from room temperature to 500 °C and strain-rate range of 10⁻³ to 10³ s⁻¹ does not exceed 10%, with an average rRMSE of 5.42% for 7075-T6 aluminum alloy and 4.09% for 65Mn high-carbon steel. The PINN model mitigates the substantial stress prediction deviations under high strain-rate and high-temperature conditions caused by the decoupling assumption of the J-C model, as well as the insufficient prediction reliability of pure ANN models for edge working conditions. This work provides an improved approach for characterizing impact mechanical behaviors of metals under wide strain-rate and temperature conditions.