Inverse problems in structural dynamics remain a significant challenge due to limited measurements, complex system behavior, and the presence of multi-scale and high-frequency responses. Physics-informed neural networks (PINNs) provide a promising framework for addressing such problems by embedding governing physical laws directly into the learning process [1]. However, practical applications have shown that the performance of standard PINN formulations is highly sensitive to how the network is designed and formulated. In this work, we systematically investigate key design choices in PINNs for inverse problems in structural dynamics from two complementary perspectives. First, we examine the role of network representation by studying the interaction between activation functions and frequency content. We show that commonly used activation functions suffer from spectral bias, which limits the ability of PINNs to capture high-frequency response components that are critical for accurate inverse identification. This limitation is alleviated by employing sinusoidal activation functions, which significantly improve frequency representation and inverse performance [2]. Second, we examine the influence of PINN architecture and problem formulation on inverse problem performance. Different architectural strategies, including sequential, parallel, and global PINN formulations, are investigated to assess how event-wise versus system-level learning affects the consistency and robustness of inverse solutions [3]. The proposed design perspectives are demonstrated through representative inverse problems in structural dynamics, including damage identification and impact-driven systems. Numerical investigations highlight that both activation function selection and PINN architectural strategy play critical roles in extending the applicability of PINNs to challenging inverse problems in structural dynamics.