SEA-PINN Enhances Physics-Informed Neural Networks with Attention Mechanism

Physics-Informed Neural Networks (PINNs) are powerful tools for solving scientific computing problems by embedding physical laws directly into the neural network training. However, their performance can sometimes be sensitive to initialization and convergence properties. Researchers introduce SEA-PINN, a novel architecture that incorporates a Squeeze-Excitation-like attention mechanism into PINNs. This mechanism dynamically recalibrates the importance of neurons across layers, leading to a highly stable initialization with nearly negligible variance and significantly reduced initial loss. SEA-PINN achieves competitive accuracy on benchmark problems, even without specialized Fourier feature embeddings, and further boosts performance when integrated with other advanced PINN variants. This demonstrates its effectiveness as a lightweight, plug-in module for enhancing nonlinear representation power, promoting robust convergence, and strengthening the reliability of physics-informed learning.