In this talk, we propose a novel Phase Field Smoothing-Physics Informed Neural Network (PFS-PINN) to solve PDEs with discontinuous coefficients. This method overcomes low solution regularity by first using a neural network to solve a phase field model, creating a smoothed approximation of the original PDE. A mixed PINN model then solves this smoothed problem, using the solution and its first-order derivatives as outputs to avoid computationally expensive second-order derivatives in the loss function. We also employ adaptive sampling to enhance performance. Numerical experiments on elliptic equations with complex microstructures confirm the accuracy and effectiveness of the PFS-PINN approach, demonstrating its potential for broader applications.