Predicting Acoustic Transmission Loss Uncertainty in Ocean Environments with Neural Networks

Round 1

Reviewer 1 Report

This paper presents a novel technique utilising neural networks to predict transmission loss variability in the form of a probability distribution function or histogram. The Neural network is trained utilising Monte-Carlo simulation (based on RAMGEO) in a wide variety of environments. Input parameters to describe the environments are sound speed, bathymetry, and seabed properties.

 

The paper is well written language-wise, the explanations that are given are clear. A few of recommendations:

 

The first results in Figure 6 merit further discussion – in particular example A shows a clear discrepancy between histograms. Is this because the assumption of a 3 parameter log normal distribution is fundamentally wrong? The TL PDF clearly transitions as a function of range. I believe the validity of LN3 over range should be discussed. Are we just seeing evidence of the central limit theorem at long range? Does it matter that at short range the LN3 is not adequate to describe the variability of TL? Would it be better to find a more suitable distribution? Is there a more suitable distribution. Is it better then to utilised the bin LN3 results instead at shorter ranges? And when can you trust the analytical LN3?

The second NN output directly provides values for each histogram bin. I assume that this does not pre-suppose any type of distribution and the NN is learning to mimic the MC histograms. If so, I would expect the NN histogram to be similar at all ranges with the MC histogram? Then why are L1 errors identical for NN-hist and NN-LN3?

The L1 errors discussions fall directly out of any discussion around the validity of the LN3, this could benefit from clarity.

Finally, the NN is effectively trained on simulated data. The paper could benefit from a discussion around the validity of the original simulation and any thoughts on validating the method with real-world data to answer the so what? question.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This article is well-motivated and well-written and the results are very interesting as they indicated the challenges and a way to use machine learning approach to construct an actual NN algorithm to predict the quite interesting Probability Density Function (PDF) of the transmission loss (TL) quantity in the uncertain ocean environment.

The authors compute with the classical Monte Carlo method a very large dataset of MC PDF TL. They use it to train several NNS. They demonstrate a quite satisfactory success rate of the constructed NNs. The proposed NNS compute very fast the PDF as compared to classical methods, the MC, among them.

The proposed scheme to deal with uncertainty is systematic. 

Question 1: The success rate is satisfactory. Do the authors believe that a factor that is responsible for this level of success  rate is the fact that the NN is trained with examples that are isolated and thus the NN does not know a priory that they stem  form a system which is continuum and its motion is governed by definite physics.  in other words, the proposed NNs are not informed by the physics underlining the datasets. Therefore, the second question is what about feeding an ensembles of examples to the NNs after reducing the typical ensemble to new reduced data, which are essential and optimal.   The reduced data from the ensemble encoded in reduced data the laws of mechanics. Perhaps this could increase the success rate of the proposed NNS. The authors might be aware of current mechanistic machine learning approaches in solid and fluid mechanics referred to as physics-informed  ANNs where the equations of motion are incorporated in ANNS algorithms as constraints.

Question 3: Could the database be replaced with examples that are produced on the basis of sensors only? Is there an analogous Monte Carlo technique of an actual physical system, for example, a local area of interest in the ocean environment. 

In summary, this article documents a systematic work with remarkable results.  One main result is the systematic treatment of the uncertainty in existing databases and in the simulation of the ocean environment for the focused purpose to explore how NN algorithms predict fast a quantity of practical interest.

 

 

  

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

This paper introduces a machine learning technology, which is used to quickly estimate PDF of TL using only a single baseline TL calculation, in order to convert the ocean environmental uncertainty into the acoustic transmission loss (TL) uncertainty. The research content has a good reference value for predicting the uncertainty of acoustic transmission loss in the ocean environment.

The specific modifications are as follows:

1. It is necessary to add an algorithm block diagram about Neural networks in the section "2.2. Methods".

2. Some syntax errors need to be corrected.

Author Response

Please see the attachment.

Author Response File: Author Response.docx