Deep Learning Framework for Solving Geoacoustic Inversion Problems using Normal Mode Theory

Abstract

Geoacoustic inversion, the process of estimating seafloor properties from acoustic measurements, is crucial for applications ranging from sonar performance prediction to environmental monitoring. However, traditional inversion methods are often computationally intensive, limiting real-time analysis and large-scale spatial characterization. This thesis presents a deep learning framework to overcome these limitations in shallow-water environments. The core approach involves training one-dimensional convolutional neural networks (1D-CNN) on large synthetic datasets generated using the KRAKEN normal mode acoustic propagation model. By learning the complex mapping from the time series of acoustic pressure to environmental parameters, these networks enable near-instantaneous geoacoustic inversion. Three main contributions are presented. First, a fully automated end-to-end system is developed for use with a single hydrophone. This system detects impulsive acoustic events, locates the source, and inverts key geoacoustic parameters, including sediment sound speed, density, and layer thickness. Validation using data from the 2022 Seabed Characterization Experiment (SBCEX) in the New England Mud Patch shows results comparable to traditional methods but achieved in milliseconds. Second, the methodology is extended to use data from distributed hydrophone arrays to estimate the spatial variability of seabed properties. By combining hundreds of individual track inversions, detailed maps of sediment sound speed and sound speed ratio across the New England Mud Patch are constructed, revealing spatial trends consistent with independent geological surveys and localized inversion studies. Third, to facilitate sensitivity analysis and potential future physics-informed machine learning, a differentiable implementation of the KRAKEN normal mode solver is developed in the Julia programming language. This enables the efficient and exact calculation of the acoustic field derivatives with respect to environmental parameters using automatic differentiation, providing valuable insights into parameter sensitivities. The results demonstrate that this deep learning framework significantly accelerates geoacoustic inversion while maintaining accuracy. It enables real-time applications and comprehensive spatial mapping previously infeasible, opening new possibilities for underwater acoustic research and bridging data-driven methods with physics-based modeling.