Acoustic field prediction in Pekeris waveguides using a learnable effective depth

Predicting acoustic fields from sparse measurements using physics-informed neural networks (PINNs) is challenging under uncertain bottom geoacoustic parameters. This paper proposes a framework incorporating a learnable effective-depth bottom boundary and two-stage training strategy, called LED-PINN. For 1 km range extrapolation beyond a 3 km measurement interval (ratio 1/3), LED-PINN reduces the mean absolute error from 7.42 dB for OceanPINN, a PINN-based model for ocean acoustic field prediction, to 2.41 dB, while improving depth extrapolation and keeping the effective-depth inversion error below 0.6 m. These results demonstrate improved acoustic field extrapolation under uncertain bottom parameters.