In this talk, I will introduce a deep learning framework for solving partial differential equations with singular solutions. The framework employs an innovative neural network architecture that incorporates prior knowledge of the problems to effectively capture the singular behavior of solutions near domain boundaries. Additionally, the framework also contains several improvements that enhance computational efficiency and accuracy of results. I will show the application of this framework to two distinct types of PDEs. Various numerical experiments will be presented to highlight its accuracy and robustness, which surpass traditional numerical methods.