Accurate prediction of time-domain electromagnetic (EM) waves is essential for designing high-frequency systems in complex environments. Traditional finite-difference time-domain (FDTD) solvers become burdensome when handling a large domain, motivating alternative approaches. Physics-informed neural networks (PINNs) offer data-efficient frameworks by embedding physical constraints, but their generalization ability is limited, often requiring retraining when source locations or material configurations change. In this work, we investigate a physics-informed deep operator network (PI-DeepONet), for modeling two-dimensional transient EM wave propagation by incorporating the time-domain Helmholtz equation. Leveraging the operator-learning structure of DeepONet, the proposed framework demonstrates enhanced generalization across diverse excitation and material conditions, including multi-source in free-space and inhomogeneous media. Each trained model predicts wave propagation and scattering for arbitrary source and scatterer configurations, including movable dielectric inclusions. The predicted spatiotemporal fields are quantitatively compared with FDTD simulations to validate accuracy and assess the model's potential as an efficient surrogate for time-domain EM analysis.