Moreover, traditional acoustic methods tend to focus on a narrow set of parameters, such as impedance or roughness, and frequently discard rich waveform data in favor of summary metrics like echo intensity or travel time. This simplification can obscure important details about sediment structure and layering. Environmental variability—such as changes in water column properties or biological interference—adds further uncertainty, making it difficult to extract reliable information. High-fidelity numerical models can help, but they are computationally intensive and impractical for real-time or large-scale applications. Inverse problems, which aim to infer sediment properties from acoustic data, are often ill-posed and sensitive to noise, leading to unstable or ambiguous results.
To address these challenges, this research proposes the use of Physics-Informed Neural Networks (PINNs) as a modern framework for seafloor characterization and classification. PINNs combine the flexibility of machine learning with the rigor of physical modeling by embedding governing equations—such as the Helmholtz equation and Biot’s poroelastic theory—directly into the neural network architecture. This allows the models to learn from data while respecting the physics of wave propagation in complex media. PINNs can handle full waveform inputs, model nonlinear and coupled effects, and generalize across diverse environments. Once trained, they serve as fast, interpretable surrogate models capable of real-time prediction and inversion. By integrating MBES and SBP data, PINNs can produce high-resolution maps of sediment properties and sub-seafloor structures, offering a scalable, accurate, and environmentally responsible solution for modern marine surveying.
Point of contact: Endrias Getachew Asgedom endrias.getachew.asgedom@kd.kongsberg.com