The inverse problem of low-temperature plasmas refers to determining discharge parameters such as voltage amplitude and frequency from plasma characteristics, including plasma density, electric field and electron temperature. Within the framework of fluid description, it is usually very challenging to address inverse problems by using traditional discretization methods. In this work, physics-informed neural networks (PINNs) are introduced to solve the inverse problem of atmospheric-pressure radio-frequency plasmas. The loss function of the PINNs is constructed by embedding three components: the main governing equations (continuity equation, Poisson equation, and drift–diffusion approximation), the discharge parameters to be inferred (voltage amplitude and frequency in this study), and additional electric field data. The well-trained PINNs can accurately recover the discharge parameters with errors within about 1%, while providing the full spatiotemporal evolution of plasma density, electric field, and flux. Furthermore, the effects of sampling positions, sampling sizes, and noise levels of the electric field data on the inversion accuracy of voltage amplitude and frequency are systematically investigated. The results demonstrate that PINNs are capable of achieving precise inversions of discharge parameters and accurate prediction of plasma characteristics under given experimental or computational data, thereby laying a foundation for the intelligent control of low-temperature plasmas.