Simulating ultrasonic wavefield propagation in solids is challenging due to their complex scattering nature. Conventional finite element method (FEM) based simulations are promising, but computationally expensive for real-time use. Physics-informed neural networks (PINNs) can accelerate ultrasonic wavefield simulations, however, most existing approaches are limited to single-mode waves and fixed scatterer (defect) locations. This limits applicability because mode conversion of ultrasonic waves is inherent in solids and a small change in defect location significantly alters their scattering behavior. To address these issues, this study proposes a novel PINN model to effectively simulate ultrasonic wavefields with mode conversions and varying defect locations. The model employs a U-net backbone to learn spatiotemporal evolution of ultrasonic wavefields, while embedding 2D elastic wave equations into the loss function to enforce physical fidelity. Model performance is evaluated to ensure visual similarity with the reference wavefields and to minimize the residuals of the wave equations. The evaluation metrics include mean absolute error, the structural similarity index and physics-based criteria. The results indicate that the proposed model can accurately predicts ultrasonic wavefields with mode conversions and varying defect locations, while providing significantly faster inference than the conventional FEM simulations.