Spectrum-adaptive physics-informed neural network for rapid ocean acoustic field prediction

To address the slow convergence of physics-informed neural networks (PINNs) in ocean acoustics, this work proposes a spectrum-adaptive physics-informed neural network (SA-PINN) built on OceanPINN for efficient modeling. SA-PINN estimates effective physical bandwidths from prior data to determine direction-dependent cutoff frequencies for normalized range and depth. This calibrates the sinusoidal representation network's maximum frequency hyperparameters, ensuring the network's spectral capacity aligns with inherent wavefield characteristics for a superior initialization. Compared with standard OceanPINN, SA-PINN achieves state-of-the-art convergence speed, significantly reducing computational costs while enhancing accuracy. Simulations and experimental data rigorously validate its effectiveness.