Abstract
In this paper, we propose a discontinuity-capturing shallow neural networkwith categorical embedding to represent piecewise smooth functions. The networkcomprises three hidden layers, a discontinuity-capturing layer, a categoricalembedding layer, and a fully-connected layer. Under such a design, we show thata piecewise smooth function, even with a large number of pieces, can beapproximated by a single neural network with high prediction accuracy. We thenleverage the proposed network model to solve anisotropic elliptic interfaceproblems. The network is trained by minimizing the mean squared error loss ofthe system. Our results show that, despite its simple and shallow structure,the proposed neural network model exhibits comparable efficiency and accuracyto traditional grid-based numerical methods.