Reconstructing the Global Stress of Marine Structures Based on Artificial-Intelligence-Generated Content

Round 1

Reviewer 1 Report

The paper is an interesting application of AI to the investigation of stress/stain states in the case of the connector structure of the offshore platform. Marginal improvements can be here suggested such as:

- some descriptive parts are maybe not so use: for instance, fig. 1, since it represents the baseline of knowledge of statistics. The same in the case of Fig. 2, representing the very very baseline for ANN.

- Figs. 3 and 8 should be better explain in the text

- Spearman correlation coefficient should be also considered, at least in one case, with the scope to try to investigate if a not linear (vs strictly linear) correlation could better explain data.

- is there any reason why ANN is preferred instead of other AI approach (e.g. Random Forest, Bayesian Methods, etc) ? if so, please, provide information in the discussion. 

- the authors' roles statement is missing.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear Authors,

The article is interesting and promising.

However, please find some suggestions to improve the scientific soundness:

-avoid acronyms in the abstract. Also, acronyms should be adequately explained in the text. Avoid the assumption that the average reader knows everything.

-line 101 and equation (2) check and correct the index n and m

-line 107 - add "Pearson correlation" reference

- all figures: add an additional sentence about why this figure seems to be important according to the title.

- section 3.1 (134-136)  rephrase

-137: index V and S - not the same font as per Fig. 2

-180-183: rephrase (write clearly) what referred to a "large/small number of elements"

- Fig 2 and Fig 8 have different fonts - all figures should have same font - see template

- why is Fig 9 important to the reader?

- in section 4.5 - Table 1: consider adding the computation power (speed) of each NN. (from Fig12,13) - in this case, Fig 12,13 could be avoided.

Also, add the used libraries.

- According to sections 4 and 5 - It is not clear if the load cases and stress data are from actual measurements or from calculated (simulated) cases. Suggesting to write clearly.

 

 

-Fig 22 - the font is too small.

- line 324 - consider avoiding the phrase "artificial intelligence",  

- 334,335 Discuss the uncertainty analysis and add a reference

- consider to add the future plans

The English language is acceptable. Minor editing of the English language required

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The reviewer has fundamental objections to the article. Only in the description of the verification experiment is the information that "the test was conducted within the elastic ranges." The lack of information about the consideration of any nonlinearity in the FE model (whether material or geometric) usually means that the FE model is linear, which has significant consequences - the different loading schemes considered in the paper (which allow the calculation of patterns for learning and testing the network) are linearly dependent! So there are not - as the author states - 3600 of them (100 different values each for 36 different angles) but only 36 (one for each angle). It means that the author prepared a so-called look-up table with all possible 36 cases calculated, after which, using the prepared tools, he selects one of these 36 cases and multiplies it by a constant to match the load value.

The author should first attach the description of the FE model. If this is a linear model, then, in the reviewer's opinion, the work is not suitable for publication. If it is a non-linear model (details are necessary), one can begin to discuss the next big problem: network overfitting. Despite attempts to explain that such a phenomenon does not occur in the case under consideration, the reviewer doubts whether this danger is avoided (even in the case of a non-linear FEM model). The overall number of patterns is 3600 (if not 36 as written above), and the network parameters probably reach hundreds of thousands (more than 30,000 output neurons - why is it necessary to obtain the stresses in EVERY FE element?) However, this discussion (and other side discussions) can be started only after clarifying the linearity of the FEM model used.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

I accept the explanation of the construction of the finite element model, the additional information provided by the authors is satisfactory.

However, I still have doubts about the learning of the network - the patterns were built by changing the value of the load and the angle of the force application. For a fixed angle, the force values changed every 1% of the maximum value (a step of 50 kN at a maximum value of 5000 kN) - this means, in my opinion, that the differences between the stress fields obtained for two adjacent load values were insignificant. This means that the differences between the learning and testing patterns were insignificant - this in turn leads to the conclusion that truthful testing or verification was not carried out at all. The differences between the learning and testing patterns may be so insignificant that the process of checking on the "testing" patterns does not fulfill the function of true testing. The network was only learned, with such a huge number of free parameters (several layers hidden in the BP network, more than 30000 outputs) there were no problems perfectly matching the patterns.

In my opinion, not 30000 stress values should have been identified, but only two: the value of the force and the angle of its inclination. And then from the pre-calculated cases (36 angles of inclination, for each 100 force values) select a ready stress field, with the load parameters closest to the identified values.

The task solved by the authors is interesting and important, but the method of solution is not acceptable in my opinion.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 3

Reviewer 3 Report

I am not fully convinced by the explanations provided by the authors. Nevertheless, some of my comments have been taken into account (new angle) - I think that the article can be (after the corrections made by the authors) submitted to the scientific community for evaluation.