1: \begin{abstract}

2: % The deep learning techniques have been introduced into solving partial differential equations (PDEs) in recent years. 

3: % Based on the universal approximation power of deep neural networks, the physics-informed neural networks (PINNs) \cite{raissi2018deep} exhibits promising results dealing with both forward and inverse PDE problems. 

4: % PDEs with a point source (mathematically expressed as Dirac delta function $\delta{(\bx)}$) are classical models in many physical problems, such as electromagnetism, heat transfer, poroelastic, and acoustics.

5: % However, the PINNs method cannot solve these singular problems directly because the Dirac delta function contains point with infinite value which cannot be computed in a computer.

6: % To address this, we present a novel approach to train a deep neural network that solves PDEs with point source singularities. 

7: % Our method tackles this problem by approximating $\delta{(\bx)}$ with a concentrated density function. 

8: % The natural signals often come with multiple frequencies, however, the standard fully-connected networks with ReLU activation are known bias towards low frequency components \cite{rahaman2019spectral}.  

9: % We improve the accuracy and convergence speed of the PINNs method by making use of the multi-scale neural network with periodic activation functions.

10: % Meantime, a new self-adaptive weighting algorithm is proposed to balance the PINNs losses between the source areas and other components.

11: % We perform experiments on several representative PDEs, including Maxwell's equations, Poisson’s equation, and Barry and Mercer’s source problem to demonstrate the effectiveness and accuracy of the proposed method.

12: % Furthermore, we exploit its robustness with various choices of hyperparameters.

13: % \end{abstract}

14: