The solution of partial differential equations (PDES) on irregular domains has long been a subject of significant research interest. In this work, we present an approach utilizing physics-informed neural networks (PINNs) to achieve irregular-to-regular domain mapping. Thus we can use finite difference method and physics-informed convolutional neural networks to solve PDEs on rectangular grids instead of the original irregular boundary. Structured grids on irregular domains are obtained by inverse mapping. We demonstrate PINN's versatile capability to produce customized structured grids tailored to diverse computational requirements, thereby significantly facilitating PDEs solving.