Abstract
We introduce a class of Sparse, Physics-based, and Interpretable NeuralNetworks (SPINN) for solving ordinary and partial differential equations. Byreinterpreting a traditional meshless representation of solutions of PDEs as aspecial sparse deep neural network, we develop a class of sparse neural networkarchitectures that are interpretable. The SPINN model we propose here serves asa seamless bridge between two extreme modeling tools for PDEs, dense neuralnetwork based methods and traditional mesh-based and mesh-free numericalmethods, thereby providing a novel means to develop a new class of hybridalgorithms that build on the best of both these viewpoints. A unique feature ofthe SPINN model we propose that distinguishes it from other neural networkbased approximations proposed earlier is that our method is both fullyinterpretable and sparse in the sense that it has much fewer connections than adense neural network of the same size. Further, we demonstrate that Fourierseries representations can be expressed as a special class of SPINN and proposegeneralized neural network analogues of Fourier representations. We illustratethe utility of the proposed method with a variety of examples involvingordinary differential equations, elliptic, parabolic, hyperbolic and nonlinearpartial differential equations, and an example in fluid dynamics.