Adaptive physics-informed neural networks for underwater acoustic field prediction

This paper introduces an adaptive physics-informed neural network for predicting underwater pressure fields. A gradient-based adaptive weighting method is proposed to address the imbalance between physics-constrained and data-fidelity terms, effectively capturing complex field structures and preserving important modal features. The origin of this imbalance is also analyzed, providing insight into the limitations of fixed-weight approaches. Validated through simulations and experimental data, this method demonstrates accurate predictions across pressure fields with varying structures and frequencies, including complex multimodal patterns. The results highlight the robustness and effectiveness of this adaptive approach, making it a promising solution for practical underwater acoustic field reconstruction.