本篇論文會介紹數值分法和類神經網路兩種方式來對偏 微分方程做數值分析。首先會以中央差分法先來做介紹，而 這是一個能數值解微分方程的傳統方法。除此之外，也會討 論針對不同類型的邊界條件所帶來的影響，分別為狄力克雷 邊界條件和諾伊曼邊界條件。接下來會介紹沈浸介面法，是 一個能處理任意維度介面問題的方法，但對於高維度問題時 沈浸介面法會遇到困難。因此，我們會利用類神經網路的方 法來嘗試能否解決此問題，且得到更好的準確度。針對介面 問題，將會使用不連續捕獲淺層神經網路法來解決。最後， 會針對連續與不連續偏的問題做降階法來看是否能得到更好 的準確度。;This thesis starts by introducing the Finite Difference Method (FDM)[1] which is one of the classic numerical methods that can solve the partial differential equations (PDEs). Moreover, we discuss the effect different types of boundary conditions on the problems, including Dirichlet and Neumann boundary conditions. After that, we will tell the method that can solve the d -dimensional interface problem which is, Immersed Interface Method (IIM)[2, 3]. Although the IIM could easily solve the interface problem, it might be difficult to deal with the high-dimensional problem. Therefore, we tried to use the Neural Network (NN) [4] to see whether it is a good way to fix it and try to get better accuracy. For the interface problem, we used the Discontinuity Capturing Shallow Neural Network (DSCNN) [5]. At last, we use a method of reduction for the continuous and discontinuous problems to find out it is a way to obtain results better than before.
