Predicting transmission loss errors through machine learning

Abstract

Uncertainty in a sound speed profile can lead to uncertainty in the associated prediction of transmission loss (TL). In order to better quantify the effect of imperfect knowledge of the sound speed profile on the acoustic propagation, a study was done comparing pairs of sound speed profiles. In each pair, one profile was treated as ground truth and the second profile was a perturbed version of the first. No single metric on the sound speed profiles, such as mixed layer depth or surface layer characteristics, correlated well with the errors in resulting TL. Several attempts at creating a more complex metric on the profile that could predict errors in the TL were also unsuccessful. However, even a rather simple machine learning approach was able to reliably predict TL errors. Results of the study will be presented and implications discussed. Uncertainty in a sound speed profile can lead to uncertainty in the associated prediction of transmission loss (TL). In order to better quantify the effect of imperfect knowledge of the sound speed profile on the acoustic propagation, a study was done comparing pairs of sound speed profiles. In each pair, one profile was treated as ground truth and the second profile was a perturbed version of the first. No single metric on the sound speed profiles, such as mixed layer depth or surface layer characteristics, correlated well with the errors in resulting TL. Several attempts at creating a more complex metric on the profile that could predict errors in the TL were also unsuccessful. However, even a rather simple machine learning approach was able to reliably predict TL errors. Results of the study will be presented and implications discussed.