On The Choice of Activation Functions in Physics-Informed Neural Network for Solving Incompressible Fluid Flows

View Video Presentation: https://doi.org/10.2514/6.2023-1803.vid This paper studies the impact of activation functions in Physics-Informed Neural Network (PINN) for solving incompressible fluid flows. A proper choice of activation function is vital so that the resulting network can efficiently solve the given incompressible fluid flow problem. This paper uses three benchmark problems: the Kovasznay flow, Taylor Green Vortex, and Beltrami flow. Six activation functions are implemented and studied: ReLU, Leaky ReLU, Sigmoid, hyperbolic tangent (tanh), Swish, and adaptive Swish function. The results show that the most accurate results are obtained using the Swish activation functions family. However, the accuracy obtained by the tanh activation function is sufficiently good, with the computational cost being smaller than that of Swish. Finally, it is observed that the ReLU activation function family yields the lowest accuracy, despite its fast training time.