Estimating the probability density function of transmission loss in an uncertain ocean using machine learning

Predicted values of transmission loss (TL) in ocean environments are sensitive to environmental uncertainties. The resulting predicted-TL uncertainty can be quantified via the probability density function (PDF) of TL. Monte Carlo methods can determine the PDF of TL but typically require thousands of field calculations, making them inappropriate for real-time applications. Thus, a variety of alternative techniques based on polynomial chaos, field shifting, modal propagation in ocean waveguides, and spatial variations of TL near the point(s) of interest have been proposed. This presentation describes an innovative approach to estimating the PDF of TL based on nominal TL, ocean environmental parameters, and machine learning. This approach has two main challenges. First, appropriate representations must be found for ground-truth PDFs of TL generated from Monte Carlo calculations so that a neural network can be constructed to predict each parameter of the estimated PDF of TL. Four such representations are considered here. And second, a framework must be developed to generate training data for the neural networks. A proposed framework that predicts candidate environments’ training utility without computing any PDFs of TL is described. The performance of this approach is analyzed and compared to that of prior techniques. [Sponsored by ONR.]Predicted values of transmission loss (TL) in ocean environments are sensitive to environmental uncertainties. The resulting predicted-TL uncertainty can be quantified via the probability density function (PDF) of TL. Monte Carlo methods can determine the PDF of TL but typically require thousands of field calculations, making them inappropriate for real-time applications. Thus, a variety of alternative techniques based on polynomial chaos, field shifting, modal propagation in ocean waveguides, and spatial variations of TL near the point(s) of interest have been proposed. This presentation describes an innovative approach to estimating the PDF of TL based on nominal TL, ocean environmental parameters, and machine learning. This approach has two main challenges. First, appropriate representations must be found for ground-truth PDFs of TL generated from Monte Carlo calculations so that a neural network can be constructed to predict each parameter of the estimated PDF of TL. Four such representations are cons...