An Improved Periodic Activation for PINNs Reconstructing Convective Flows

This paper investigates neural network architectures using complex exponential functions as periodic activations, which generate paired sine and cosine outputs. The researchers evaluate this approach against comparable sine-activated multi-layer perceptrons specifically for the task of reconstructing convective flows. Results demonstrate the effectiveness of the proposed periodic activation strategy in enhancing physics-informed neural network performance for fluid dynamics applications.