Predicting the Porosity in Selective Laser Melting Parts Using Hybrid Regression Convolutional Neural Network
Round 1
Reviewer 1 Report
The manuscript applied RCNN algorithm for porosity prediction and optimized by Bee Algorithm to improve accuracy. Compared to traditional image binarization, the machine learning model shows obvious improvement in prediction accuracy without manual determination once the model has been trained. The manuscript is written more in a technical report style. Some details are not quite necessary and could be distracting the major research merits.
Major Comments about the proposed ML method:
1. It is unclear why select image binarization as comparison - is this one of the most widely used methods? Or is there any other advanced method used for porosity prediction?
2. Section 3.7 - how the average SSIM for 3000 samples is calculated? Is there a one-to-one correspondence between the actual image and the artificial image?
3. Table 8 - is the first version or final version?
4. Table 10 – generally, the Neural Network model performance is evaluated only based on the test set. In addition, the training/validation accuracy only shows the training performance and is used for model optimization. However, the average accuracy of the training/validation/test dataset is not meaningful as an indication for model assessment.
5. The experiments clearly show the prediction results, but the specific RCNN architecture framework is not given.
Author Response
Cover Letter for Reviewer 1
- Reviewer Comment: It is unclear why select image binarization as comparison - is this one of the most widely used methods? Or is there any other advanced method used for porosity prediction?
Response: Image binarization is one of the two main existing methods used in porosity estimation along with the Archimedes method, but the Archimedes method might be used in the case of producing real SLM parts. In this paper, artificial porosity images were created, so RCNN results were compared with the image binarization method. This clarification was added at the beginning of section 4.
- Reviewer Comment: Section 3.7 - how the average SSIM for 3000 samples is calculated? Is there a one-to-one correspondence between the actual image and the artificial image?
Response: Samples of 100 slices of real images were taken and compared with every 100 slices for each cube, so the first real image is compared with the first slice of the cube and the second real image is compared with the second slice and so on until slice 100. The same process is repeated for each of the 30 cubes, and then the average was calculated for 3000 values of SSI.
- Reviewer Comment: Table 8 - is the first version or final version?
Response: Table 8 (Table 9 in the revised manuscript) shows the results for the final version of artificial porosity images.
- Reviewer Comment: Table 10 – generally, the Neural Network model performance is evaluated only based on the test set. In addition, the training/validation accuracy only shows the training performance and is used for model optimization. However, the average accuracy of the training/validation/test dataset is not meaningful as an indication for model assessment.
Response: The CNN assessment changed to be based on the testing set accuracy.
- Reviewer Comment: he experiments clearly show the prediction results, but the specific RCNN architecture framework is not given.
Response: The architecture is described in detail in section 4.2.1.
Attached is the revised manuscript
Author Response File: 
Author Response.docx
Reviewer 2 Report
The authors did a very actual research about using artificial intelligence for the porosity prediction in the laser powder bed fusion process. Such an approach is very important, especially for the solutions dedicated for aircraft and space industry, where quality plays the most important role. Unfortunately, this manuscript, in a present form suffers from many imperfections which I listed below:
1. Please use proper nomenclature related to ISO/ASTM 52900 standard terminology. SLM is a commercial name - you can use it but first you need to put the standarized technology description. The same issue is with Additive Layer Manufacturing - it is also standarized. Please correct it.
2. There is a lack of proper literature review. The authors should properly analyse the actual state of the art, especially that using AI is very popular in the AM. Based on such a review the authors should highlight the novelty of their research.
2a. What is original and new in your research? It is not clear in a present form.
3. Figure 1 - if you took this image put proper information about the persmission of usage - you have it described in the guide for authors.
4. I do not understand the aim of the chapter 2. The Laser Powder Bed Fusion is a very complex process which should be deeply undestand and described by the authors. Instaed of application description please pu proper description of the process including thermodynalical phenomena description which have the most important influence on the porosity generation.
5. There is a lack of an experiment descritpion - used apparatus and software.
6. The porosity which you have analysed is related to gas porosity which is not directly connected with the exposure velocity or laser power but with the powder humidity (which was probably too high before the process). I am very sorry but it seems that you analysed wrong type of pores.
7. The authors focused mostly on the AI but to do it properly there is a need of conduct proper experimental part of the research which is described here in a very low quality - you have to improve it.
8. The discussion part has to be improved. The authors should describe their outcomes point-by-point with using quantified values.
To make this manuscript possible to publish the authors must put a significant improvements. In a present form I suggest to reject this manuscript and at the same time I encourage authors for the resubmission after they put proper improvements.
Author Response
Cover Letter for Reviewer 2
- Reviewer Comment: Please use proper nomenclature related to ISO/ASTM 52900 standard terminology. SLM is a commercial name - you can use it but first you need to put the standarized technology description. The same issue is with Additive Layer Manufacturing - it is also standarized. Please correct it.
Response: The standard terminology for ISO/ASTM 52900 was used at the beginning of sections 1 and 2.
- Reviewer Comment: There is a lack of proper literature review. The authors should properly analyse the actual state of the art, especially that using AI is very popular in the AM. Based on such a review the authors should highlight the novelty of their research.
Response: state-of-the-art studies are shown in section 2.5, the section ended with the gap in measuring the porosity in the existing method and linked with the contribution of the paper.
2a. Reviewer Comment: What is original and new in your research? It is not clear in a present form.
2a. Response: The contribution of the paper is developing a new approach based on the use of a Regression Convolutional Neural Network (RCNN) algorithm to predict the percent of porosity in CT scans of finished SLM parts, without the need for subjective difficult thresholding determination to convert a single slice to a binary image. In order to test the algorithm, as the training of the RCNN would require a large amount of experimental data, artificial porosity images mimicking real CT scan slices of the finished SLM part were created.
- Reviewer Comment: Figure 1 - if you took this image put proper information about the persmission of usage - you have it described in the guide for authors.
Response: The reference number is shown in the caption and introduced just before the figure.
- Reviewer Comment: I do not understand the aim of the chapter 2. The Laser Powder Bed Fusion is a very complex process which should be deeply undestand and described by the authors. Instaed of application description please pu proper description of the process including thermodynalical phenomena description which have the most important influence on the porosity generation.
Response: Section 2 presents a review of the powder bed fusion process, showing the way of working, thermodynamical phenomena (added in section 2.2), parameters, open issues, and state-of-the-art studies.
- Reviewer Comment: There is a lack of an experiment descritpion - used apparatus and software.
Response: Section 3 starts with a flow chart showing the steps of creating artificial porosity images, and then it is followed by subsections (3.1, 3.2, 3.3, 3.4, 3.5, 3.6, and 3.7) explaining each step in detail showing the software used for each task. The section ended with section 3.8 showing study limitation and stating that the study is not aimed to study the porosities in depth, so no experiments have been conducted, but the paper proposes a method that will enhance such studies, particularly in predicting the percent of porosity as demonstrated in the results of section 4. In the future, real experiments can be conducted to produce real porosity images.
- Reviewer Comment: The porosity which you have analysed is related to gas porosity which is not directly connected with the exposure velocity or laser power but with the powder humidity (which was probably too high before the process). I am very sorry but it seems that you analysed wrong type of pores.
Response: The porosity analysed is the keyhole porosity as the data used to establish the two equations in section 3.1 are related to keyhole porosity that arises from energy input excess, which is a result of increasing the laser power or decreasing the scanning speed (increasing energy density), it is caused when trapping bubbles of vapor within the melt pool. The created artificial porosity images in section 3 were used only to test the proposed CNN algorithms developed in the section 4. But we acknowledge that further work is required to tune our approach for the creation of artificial images mimicking specifically a wider range of porosity types. This was added to our conclusion. The study is not aimed to study the porosities in depth, so no experiments have been conducted, but the paper proposes a method that will enhance such studies, particularly in predicting the percent of porosity as demonstrated in the results of section 4. In the future, real experiments can be conducted to produce real porosity images.
- Reviewer Comment: The authors focused mostly on the AI but to do it properly there is a need of conduct proper experimental part of the research which is described here in a very low quality - you have to improve it.
Response: The description of the steps for creating artificial porosity images is enhanced in section 3.
- Reviewer Comment: The discussion part has to be improved. The authors should describe their outcomes point-by-point with using quantified values
Response: The discussion is enhanced by highlighting the contribution of the paper with quantitative measures at the end of section 4.
Attached is the revised manuscript.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
All minor issues have been corrected properly but unfortunately, the most important faults are still present in this manuscript. Dear authors, please look at this paper which would clarify the difference between gas porosity and keyhole porosity:
DOI: 10.1016/j.jmst.2020.03.070
Maybe you should try to make a cross-section in the perpendicular area?
In my opinion, subchapter 2.2 should be described in a more detailed way.
I suggest additional research or deeper microstructural analysis of the samples because, in the present form, the reader who is into the AM would not be convinced about the porosity type. Unfortunately, in my opinion, such faults disqualify this manuscript for publication. I am sorry but I will maintain my decision and suggest the rejection. Additionally, I suggest extending your research, correcting it, and resubmitting your manuscript once again.
Author Response
Dear,
Section 2.2 was enhanced with a more detailed explanation, and additional research about porosity types was added in sections 2.4.1, 2.4.2, and 2.4.3 and then it is linked with artificial porosity images creation in sections 3.3 and 3.4.
Attached is the revised manuscript.
Thank You
Author Response File: Author Response.docx
Round 3
Reviewer 2 Report
Dear Authors,
After all corrections, this manuscript could be published. You have to clarify the usage of images from others publications (you will be asked to do it during the authors' proof - reading.
