The accurately and rapidly delineating surface contours is crucial for ultrasonic immersion nondestructive testing (NDT) of complex curved-surface components. However, directly employing ultrasonic signals for interface reconstruction in dual-layer media remains a challenge. In this article, a physics-informed neural network (PINN) is developed for the reconstruction of the interface between water and unknown surface components. By incorporating the nonlinear equations of Fermat’s principle as additional constraints in the loss function, a neural network is constructed that is dual driven by both data and physical information. The simulation and experimental results show that the maximum error (ME) of the interface reconstruction by this method is below 0.5 mm, the average relative error (ARE) is less than 1.0%, and the computation time is less than 1 s. The imaging results of V-shaped crack defects using different interface reconstruction methods are further compared, verifying the significant advantages of PINN in ultrasonic array inspection of complex curved components.