In this paper, we investigate a recently proposed neural network method for learning the input-output behavior of power system devices using field or simulation data. This method adapts Neural Ordinary Differential Equations to represent components described by differential-algebraic equations, which are essential in power system component modeling. The approach yields continuous-time state-space models that can be seamlessly integrated into existing transient stability simulation tools. We utilize this method to train surrogate models using data obtained from transient stability simulations of a standard inverter model. We evaluate the performance of these surrogate models and analyze how their accuracy depends on neural network architecture and training parameters, such as the volume of required data. Furthermore, we assess the model’s ability to represent various transient behaviors, including responses to active and reactive power commands and low-voltage ride-through events.