Geoacoustic parameter estimation using machine learning techniques

The Airy phase region corresponding the minimum group velocity of the dispersive acoustic normal modes in shallow water are extremely sensitive to bottom properties. The group speed minima and the associated frequency data for two lower order modes were used to train a neural network to predict the bottom parameters in a previous study (Potty et al., 2019). The geoacoustic model for the bottom consisted of a sediment layer over acoustic basement. Synthetic data generated by varying the values of the sound speeds in the sediment layer and basement were used to train and evaluate the technique. The output of the neural network was used as the background model for a linear perturbation inversion. Data collected from the Shelf break Primer experiment were used to test the algorithm and the preliminary results were promising. This study will continue the previous work by incorporating higher order modes in the training data, testing with data from other experiments and using other machine learning techniques. The performance of this hybrid approach (neural network combined with linearized inversion) will be compared with fully non-linear inversion, both in terms of accuracy and computation time. [Work supported by the Office of Naval Research.]
    • Correction
    • Source
    • Cite
    • Save