MEMS High Aspect Ratio Trench Three-Dimensional Measurement Using Through-Focus Scanning Optical Microscopy and Deep Learning Method

Round 1

Reviewer 1 Report

The report presented by the authors on 'MEMS high aspect ratio trench three-dimensional measurement using TSOM and deep learning method' scientifically interested. It can be considered for publication, if the authors are ready to consider the following suggestions:

(1) Grammartical structure should be enhanced throughout the manuscript. For instance, discussion of predicted results should be reported in past tense, this among others should be carefully looked into.

(2) Authours should avoid using abbreviation in the title.

(3) Authors should avoid repeatation of title words in keywords as much as possible.

(4) Discussion of the predicted results should be written comprehensively.

Author Response

The report presented by the authors on 'MEMS high aspect ratio trench three-dimensional measurement using TSOM and deep learning method' scientifically interested. It can be considered for publication, if the authors are ready to consider the following suggestions:

(1)    Grammartical structure should be enhanced throughout the manuscript. For instance, discussion of predicted results should be reported in past tense, this among others should be carefully looked into.

Thanks for your suggestion.

Grammartical structure has been checked, and some mistakes have been corrected in the revised version.

(2)    Authours should avoid using abbreviation in the title.

Thanks for your suggestion.

The “TSOM” in the title has been replaced by “through-focus scanning optical microscopy” in the revised version. MEMS is a widely used abbreviation, so it’s kept in the title.

(3)    Authors should avoid repeatation of title words in keywords as much as possible.

Thanks for your suggestion.

The keywords have been replaced by other appropriate words in the revised version.

(4)    Discussion of the predicted results should be written comprehensively.

Thanks for your suggestion.

We add some discussion and analysis in the revised version.

Reviewer 2 Report

The paper uses the DL method to interpret TSOM images of MEMS  HAR structures to measure depth and width of the trenches in the structure. The paper is well written; however, since its main contribution is the  use  of the CNN  and  the model presented in Fig.3, presenting some details about the  extraction of  the   kernels' values improves  the paper.  For example it should be noted that in the training procedure  which data in the figure is set to the desired values.  Or, the 3000 training images belong to how many trenches.  Also a brief discussion on how to reach the proposed model and number of the layers would be useful. 

- In the first line of the Results section the SVR model should be corrected.

 - A newly published  paper by the authors in Chinese Optics  that uses ML method for TSOM images should be referred in the paper.

Author Response

1. The paper uses the DL method to interpret TSOM images of MEMS HAR structures to measure depth and width of the trenches in the structure. The paper is well written; however, since its main contribution is the use of the CNN and the model presented in Fig.3, presenting some details about the extraction of the kernels' values improves the paper. For example it should be noted that in the training procedure which data in the figure is set to the desired values.  Or, the 3000 training images belong to how many trenches. Also a brief discussion on how to reach the proposed model and number of the layers would be useful. 

Thanks for your suggestion.

More details are added to the Fig. 6 and corresponding demonstration in the revised version. The proposed model and number of the layers are reached by experience and comparative experiments. We didn’t give a discussion about it because it is difficult to describe in short discussion. I’m sorry for that.

2. In the first line of the Results section the SVR model should be corrected.

Thanks for the remind.

The mistake has been corrected in the revised version.

3. A newly published paper by the authors in Chinese Optics that uses ML method for TSOM images should be referred in the paper.

Thanks for your suggestion.

The newly published paper has been added as reference in the revised version.

Reviewer 3 Report

- Please use vectorized format for all figures. Currently, all images are rasterized which reflects poor scientific practice.

Author Response

  1. Please use vectorized format for all figures. Currently, all images are rasterized which reflects poor scientific practice.

Thanks for your suggestion.

The quality of the figures has been improved in the revised version, and the original figure files are provided to the editor.

We also optimize the structure of method part in the revised version.

Reviewer 4 Report

In this paper, the author proposed a method to measure the width and depth of high aspect ration trenches using through-focus scanning optical microscopy and deep learning. It might attract interest from the reader but some issues should be addressed first.

1. The quality of the figures should be improved.

2. More details on traditional machine learning method should be presented in Methods section.

3. In line 199, the author declared that "(ML predicted results)are much lower than DL predicted results". It is contrary to fact in Figure 5 that DL predicted results are much lower.

Author Response

In this paper, the author proposed a method to measure the width and depth of high aspect ration trenches using through-focus scanning optical microscopy and deep learning. It might attract interest from the reader but some issues should be addressed first.

1. The quality of the figures should be improved.

Thanks for your suggestion.

The quality of the figures has been improved in the revised version, and the original figure files are provided to the editor.

2. More details on traditional machine learning method should be presented in Methods section.

Thanks for your suggestion.

We add a reference about the traditional machine learning method in the revised version.

3. In line 199, the author declared that "(ML predicted results) are much lower than DL predicted results". It is contrary to fact in Figure 5 that DL predicted results are much lower.

Thanks for the remind.

The mistake has been corrected in the revised version.