AE-PINNs: Attention-enhanced physics-informed neural networks for solving elliptic interface problems (2506.18332v1)

Abstract: Inspired by the attention mechanism, we develop an attention-enhanced physics-informed neural networks (AE-PINNs) for solving elliptic interface equations. In AE-PINNs, we decompose the solution into two complementary components: a continuous component and a component with discontinuities across the interface. The continuous component is approximated by a fully connected neural network in the whole domain, while the discontinuous component is approximated by an interface-attention neural network in each subdomain separated by the interface. The interface-attention neural network adopts a network structure similar to the attention mechanism to focus on the interface, with its key extension is to introduce a neural network that transmits interface information. Some numerical experiments have confirmed the effectiveness of the AE-PINNs, demonstrating higher accuracy compared with PINNs, I-PINNs and M-PINNs.