This study develops a physics-informed neural network (PINN) model to efficiently generate high-fidelity local circulation patterns in Taiwan while preserving an accurate representation of the physical relationship between generated local turbulence and upstream synoptic flow regimes. Large ensemble semi-realistic simulations were conducted using a high-resolution (2 km) TaiwanVVM model, in which the critical characteristics of the various synoptic flow regimes are carefully selected to concentrate on the effects of local circulation variations. A variational autoencoder (VAE) was constructed to capture essential representations of local circulation scenarios associated with the lee vortices by training on the ensemble dataset. The VAE’s latent space effectively encapsulated synoptic flow regimes as controlling factors, aligning with the physical understanding of Taiwan’s local circulation. The critical transition of flow regimes under the southeasterly synoptic flow regimes is also well represented in the VAE’s latent space, indicating that the VAE can learn the nonlinear characteristics of the multiscale interaction of the lee vortex. We further construct a reduced-order model for local circulation predictions under diverse synoptic conditions. This physics-informed VAE ensures the accurate prediction of the nonlinear characteristics of multiscale interactions between synoptic flows and the turbulence induced by topography, thereby accelerating the assessment of the local circulation responses under various climate change scenarios.