Density Prediction in Powder Bed Fusion Additive Manufacturing: Machine Learning-Based Techniques

Round 1

Reviewer 1 Report

Thank you for performed work. Please find attached the file with my comments and suggestions. I hope you find them useful.

Comments for author File: Comments.pdf

Author Response

Responses to the reviewer's comments

Comment 1: Line 21: Specify that SS316L is stainless steel so that readers quickly understand the material used for the PBF-AM.

Response:  Thank you for the comment. SS316L is replaced with stainless steel (SS) 316L in the revised manuscript.

Comment 2: Lines 24-26: Are conclusions relevant to the performed work?

Response: Yes, the developed ML model can be a new direction to predict the process parameters and also can be helpful in future applications.

Comment 3: The sentence “Machine learning (ML) is one of the artificial intelligence tool which uses past data to learn the relationship between input and output and helps to predict the future trends” from the Abstract appears about 5 times in the text in the rephrased but very similar appearance. The reviewer recommends minimizing reusing similar statements so many times by using them only where it is necessary to remind a reader or by improving the structure of the manuscript so that reader’s attention is not distrusted.

Response: Thank you for the comment. In response to the reviewer’s comment, sentence repetition has been checked and avoided in the revised manuscript.   

Comment 4: Lines 419-420: It might worth specifying the accuracy of 97.5% for ANN

Response: Thank you for the comment, in the updated manuscript the sentence rephased as follows: “The second-best model was found to be the ANN, which provided the highest correlation values i.e. 95.5%.”

Comment 5: Finally, language and grammar needs a little more work. As example from the first paragraph of Introduction:

• Line 31: Double-check if you want “Additive” to be capitalized because it was not in the Abstract and further in the manuscript.
• Lines 31-32: Does “…such as complexity-free, customization…” misses a word in-between? • Lines 32-33: “…properties, … parts, … single-piece part..” is inconsistent.
• Lines 31-33: “capabilities…such as…attracts” is inconsistent.
• Line 46: “Required” incorrect tense.
• Lines 49-50: “Machine Learning (ML)” and “Machine learning” capital letters are inconsistent.
• Lines 69-70: “Hence in the AM part density optimization is almost important to measure the quality of the building part” sentence requires correction.
• Line 266: “a different ML technique”. The reviewer believes the authors meant “different ML techniques”. If the authors meant the ANN model they discuss in Results, it needs to be presented in Methodology.
• Line 284: “how” is extra

Response: Thank you for your valuable time and suggestions. All the mentioned corrections related to English language and grammar have been incorporated in the revised manuscript and highlighted with Green colour.

Reviewer 2 Report

The manuscript describes the machine learning techniques optimized for powder bed fusion additive manufacturing. The author selected methods such as ANN, KNN, SVM, and LM to minimize errors and mentioned which method was most appropriate. However, it is not recommended to be published for the following reasons.

Comment 1:

Throughout the paper, there is insufficient information on the real SS316L used as a comparison target, so the reliability of the real data is poor. Also, it is questionable whether calculating from just one real data can produce reasonable results. The source of data and sample information such as atomic ratio weight percent should be provided for reliable research data.

Comment 2:

There is a test dataset for each machine learning model, but no description for it. The author should add information for each dataset.

Comment 3:

Typos and grammatical errors are frequently found throughout the manuscript. It is recommended that the author take English proofreading.

Comment 4:

The forms and numbering of each figure are all different. Some are thick lines around and some are missing. It would be better to unify the form.

Author Response

Responses to the reviewer's comments

Comment 1: Throughout the paper, there is insufficient information on the real SS316L used as a comparison target, so the reliability of the real data is poor. Also, it is questionable whether calculating from just one real data can produce reasonable results. The source of data and sample information such as atomic ratio weight percent should be provided for reliable research data.

Response:  Thanks a lot for your comment. All the data collected from the literature work is on standard SS316L material. The chemical composition is also standard for SS316L ((wt.%) 16–18% Cr, 10–14% Ni, 2– 3% Mo, b0.03% C, b1% Si, b2% Mn, b0.045% P, b0.03% S, b0.1%N, and Fe balance). In this work, we utilized available data of PBF-AM of SS316L and developed a density prediction model. Since very limited experimental work reported related to process parameter variation for PBF-AM of SS316L. The data collected are less due to growing research, this aspect has been included in the limitation of the study. As the research advances, future studies could be done for improving the predictability of process parameters.

Comment 2: There is a test dataset for each machine learning model, but no description for it. The author should add information for each dataset.

Response: Thank you for your comment. The data set is provided in the Appendix section Table A1 and relevant information has been included in the revised manuscript.

Comment 3: Typos and grammatical errors are frequently found throughout the manuscript. It is recommended that the author take English proofreading.

Response: Thank you for your comment. The manuscript has been thoroughly checked for typos and grammatical errors and corrected. Moreover, English proofreading has been done to ensure the proper sentence flow.  

Comment 4: The forms and numbering of each figure are all different. Some are thick lines around and some are missing. It would be better to unify the form.

Response: Thank you for the comment. All the figures have been updated in the revised manuscript as per the reviewer suggestion.

Reviewer 3 Report

A model for predicting the partial density of SS316L is developed, and the model is evaluated and analyzed by R square value and different error functions. The methodology is the application of four machine learning methods (ANN, KNN, SVM, LR).

1. Please review and improve the English language and grammar of this manuscript.

2. The title of the literature review should be a secondary title. The introduction does not clearly point out what the existing research contribution of this paper is, but only applies four ml technologies to density prediction, which is not convincing.

3. The writing level of the thesis needs to be improved. Many sentences are difficult to understand. In addition, it is suggested to simplify the article and delete unnecessary redundancy.

For example, (1) in the introduction, “Hence in the AM part density optimization is almost important to measure the quality of the building part.”

(2) in the introduction, “The data was first classified and filtered according to input parameters (Laser Power, Layer thickness, Hatch spacing, and scanning speed) versus output as the density of SS316L.”

4. The introduction of the four machine learning models is too long, and the principle content is less. It is better to use vector diagram for schematic diagram.

5. (B) in Figure 7 is not clearly expressed and lacks necessary legend and text explanation; What is the Y equation? How to divide the data set and get the result graph, lack of text description; Black thick wireframe and shadow can be removed; Change to vector diagram; B there is an extra thin gray line on the right in the figure, which is meaningless.

6. Figure 9, figure 10 and Figure 12: modify the up-down alignment of the figure, add the legend, and remove the gray thin lines. Coordinate axis labels are changed to be uniformly bold or not bold; Figure 11 lacks necessary text explanation, so it is of little significance to place this figure at this position; Fig. 13 abscissa labels have different formats. The first three are abbreviations and the last one is the full name; The values in the blue histogram in Figure 14 are not clear, so a different color can be used.

7. Among the four machine learning model methods, the number of test sets used is different, there are three and four, so it is impossible to use a unified standard to evaluate the quality of the model.

8. The discussion did not mention the superiority of the model established by itself, and did not explain why SVM is the most suitable for predicting the density in PBF. Only figure 13 is not convincing enough, so relevant comparative literature can be added.

9. In the conclusion part, the accuracy of SVM model in predicting density is 96.01%, but in the paper, there are only 3 or 4 test sets in the form of histogram, which can not clearly express the comparison between the prediction of the model and the actual results.

Author Response

Responses to the reviewer's comments

Comment 1: Please review and improve the English language and grammar of this manuscript.

Response: Thank you for your comment. The manuscript has been thoroughly checked for typos and grammatical errors and corrected. Moreover, English proofreading has been done to ensure the proper sentence flow.  

Comment 2: The title of the literature review should be a secondary title. The introduction does not point out what the existing research contribution of this paper is but only applies four ml technologies to density prediction, which is not convincing.

Response: Thank you for your comment. In response to the reviewer’s comment, the literature review title has been made as a secondary title in the revised manuscript (Please see line no. 82).

Comment 3: The writing level of the thesis needs to be improved. Many sentences are difficult to understand. In addition, it is suggested to simplify the article and delete unnecessary redundancy.    

For example, (1) in the introduction, “Hence in the AM part density optimization is almost important to measure the quality of the building part.”

(2) in the introduction, “The data was first classified and filtered according to input parameters (Laser Power, Layer thickness, Hatch spacing, and scanning speed) versus output as the density of SS316L.”

Response:  Thank you for your valuable comments. The manuscript has been thoroughly checked for difficult-to-read sentences and corrected. Moreover, English proofreading has been done to improve the writing level of the manuscript. In response to the reviewer’s comment, the article has been simplified, and unnecessary redundancy has been deleted. Also, the mentioned corrections have been incorporated in the introduction section of the revised manuscript.    

Comment 4:  The introduction of the four machine learning models is too long, and the principle content is less. It is better to use vector diagram for schematic diagram.

Response: Thank you for your valuable comment. The introduction of ML models has been discussed in detail to better understand the concept to the readers. Moreover, the study depicts the application of ML in material science, and thus such discussion would help researchers further adopt such techniques in executing other research areas. The vector diagrams (figure 3, 5 and 6) have already been included to explain ML models.    

Comment 5: (B) in Figure 7 is not clearly expressed and lacks necessary legend and text explanation; What is the Y equation? How to divide the data set and get the result graph, lack of text description; Black thick wireframe and shadow can be removed; Change to vector diagram; B there is an extra thin gray line on the right in the figure, which is meaningless.

Response: Thank you for your comment. The legend has been updated in all the similar figures with detailed explanations in the revised manuscript. The description of equation y is also added in the section as highlighted in yellow colour. Also, the shadow effect has been removed from the figures as suggested. 

Comment 6 : Figure 9, figure 10 and Figure 12: modify the up-down alignment of the figure, add the legend, and remove the gray thin lines. Coordinate axis labels are changed to be uniformly bold or not bold; Figure 11 lacks necessary text explanation, so it is of little significance to place this figure at this position; Fig. 13 abscissa labels have different formats. The first three are abbreviations and the last one is the full name; The values in the blue histogram in Figure 14 are not clear, so a different color can be used.

Response: Thank you for your valuable comment. All the figures are modified into a uniform format in revised manuscript. Figure 13 is also updated with the uniform labels.

Comment 7 : Among the four machine learning model methods, the number of test sets used is different, there are three and four, so it is impossible to use a unified standard to evaluate the quality of the model.

Response: Thank you for the comment. Based on the ML theory, the train and test data have been divided into 70% and 30%, respectively, for four models.

Comment 8 : The discussion did not mention the superiority of the model established by itself, and did not explain why SVM is the most suitable for predicting the density in PBF. Only figure 13 is not convincing enough, so relevant comparative literature can be added.

Response: Thank you for the comment. In the discussion section, each model is compared with the other with detailed information. The similar reported literature is also included in the discussion section from line 359-362 in the revised manuscript.

Comment 9 :  In the conclusion part, the accuracy of SVM model in predicting density is 96.01%, but in the paper, there are only 3 or 4 test sets in the form of histogram, which can not clearly express the comparison between the prediction of the model and the actual results.

Response: Thank you for the comment. The comparison between actual and predicted data is plotted to obtain the best fit line between predicted and actual values. In this work, we utilized available data of PBF-AM of SS316L and developed a density prediction model. Due to the limited amount of data the training and testing data is divided into 70% for testing and 30% for training. In the near future, this model can also help carry forward similar studies with large data sets and for the other mechanical properties prediction.

Round 2

Reviewer 1 Report

Dear authors,

Thank you for proofreading the manuscript and addressing some of the suggestions. English and reviewer's specific comments are addressed sufficiently. However, the reviewer believes that two things still need to be addressed, as discussed in the general comment of the first reviewer's response:

1. Improving manuscript structure and Introduction. As mentioned before, some parts of the Methodology and the Result present many new literature references that are not shown in the Introduction. It is essential that the Introduction is comprehensive. To make it concise and easy to read for researchers that will learn from your study, tables with logically structured data and only important descriptions are preferred.

2. The amount of performed work in Results is limited. As in the first reviewer's response, it is recommended to extend the analysis by utilizing the model to a different set of additively manufactured parts or in another way that the authors consider appropriate.

Author Response

Dear Reviewer,

We would like to thank you for your review. Below you can find our answers to your comments:: 

Comment 1. Improving manuscript structure and Introduction. As mentioned before, some parts of the Methodology and the Result present many new literature references that are not shown in the Introduction. It is essential that the Introduction is comprehensive. To make it concise and easy to read for researchers that will learn from your study, tables with logically structured data and only important descriptions are preferred.

Response:  Thank you for your comment. The manuscript introduction has been enhanced by including recent literature related to methodology and findings. Following revision has been done along with inclusion of recent research articles [8-19], [25-27].

“The present study explores the application of ML techniques such as ANN, KNN, SVM and linear regression to predict the density of additively manufactured parts. ANN technique utilizes a principle similar to human intelligence [13]. Further, it establishes the relationship between input and output parameters using past experience and trends. It contains numerous neurons and hidden layers associated to parameter complexity [14]. K-nearest neighbor (KNN) is used for classification and regression analysis and does not store any learning outcome from the training dataset [15]. Support Vector Machine (SVM) is one of the most promising supervised machine learning algorithms applied for regression and classification problems [16]. The linear regression model is simple to perform and easy to interpret the output analysis. However, linear regression assumes a linear relationship between the parameters. Linear regression also looks at a relationship between the mean of dependent variables and independent variables. Linear regression is not a complete description of relationships among variables. The data-driven ML model was developed in this work to predict the density of SS316L built by powder-based AM. SS316L has extensive applications in the Bio-Implant where the complexity and mechanical property must be precisely taken care [17]. The density is directly related to all other mechanical properties. The higher density would improve tensile strength, hardness, and other mechanical properties [18]. Controlling density is crucial in the powder-based AM due to several parameters at deposition [19]. The layer-wise deposition in AM is prone to more porous parts than in other manufacturing processes. Hence, density optimization is significant in measuring the quality of the AM-built part. The selection of process parameters concerning optimum density will help predict the quality of the part and the mechanical properties.

Several researchers have attempted to apply ML approaches to predict the output with respect to different manufacturing processes. For instance, density prediction was attempted using ML for flash sintering process [25], porosity of the building part was predicted using physics-based ML techniques showing good performance with a minimum error rate of 10-25% [26] and random forest technique was used to predict the density of build part [27]. Despite the different ML applications for additive manufacturing, the density prediction of powder-based AM of SS316L has not yet been reported. This study covers a novel approach to predict the density of additively manufactured SS316L parts. The developed models are also compared to select the best model for predicting density.”

8. Wang, C., Tan, X.P., Tor, S.B., Lim, C.S.: Machine learning in additive manufacturing: State-of-the-art and perspectives, (2020). https://doi.org/10.1016/j.addma.2020.101538.
9. Gong, X., Zeng, D., Groeneveld-Meijer, W., Manogharan, G.: Additive manufacturing: A machine learning model of process-structure-property linkages for machining behavior of Ti-6Al-4V. Mater. Sci. Addit. Manuf. 1, 6 (2022). https://doi.org/10.18063/msam.v1i1.6.
10. Smoqi, Z., Gaikwad, A., Bevans, B., Kobir, M.H., Craig, J., Abul-Haj, A., Peralta, A., Rao, P.: Monitoring and prediction of porosity in laser powder bed fusion using physics-informed meltpool signatures and machine learning. J. Mater. Process. Technol. 304, 117550 (2022). https://doi.org/10.1016/j.jmatprotec.2022.117550.
11. Snow, Z., Diehl, B., Reutzel, E.W., Nassar, A.: Toward in-situ flaw detection in laser powder bed fusion additive manufacturing through layerwise imagery and machine learning. J. Manuf. Syst. 59, 12–26 (2021). https://doi.org/10.1016/j.jmsy.2021.01.008.
12. Liu, R., Liu, S., Zhang, X.: A physics-informed machine learning model for porosity analysis in laser powder bed fusion additive manufacturing. Int. J. Adv. Manuf. Technol. 113, 1943–1958 (2021). https://doi.org/10.1007/s00170-021-06640-3.
13. Rathi, N.K., Rathi, N.: An application of ANN for modeling and optimisation of process parameters of manufacturing process: A review. Int. J. Eng. Appl. Sci. Technol. 04, 127–134 (2020). https://doi.org/10.33564/ijeast.2020.v04i12.017.
14. Stangierski, J., Weiss, D., Kaczmarek, A.: Multiple regression models and Artificial Neural Network (ANN) as prediction tools of changes in overall quality during the storage of spreadable processed Gouda cheese. Eur. Food Res. Technol. 245, 2539–2547 (2019). https://doi.org/10.1007/s00217-019-03369-y.
15. Abu Alfeilat, H.A., Hassanat, A.B.A., Lasassmeh, O., Tarawneh, A.S., Alhasanat, M.B., Eyal Salman, H.S., Prasath, V.B.S.: Effects of Distance Measure Choice on K-Nearest Neighbor Classifier Performance: A Review, (2019). https://doi.org/10.1089/big.2018.0175.
16. Wu, D., Wei, Y., Terpenny, J.: Predictive modelling of surface roughness in fused deposition modelling using data fusion. Int. J. Prod. Res. 57, 3992–4006 (2019). https://doi.org/10.1080/00207543.2018.1505058.
17. AlFaify, A., Hughes, J., Ridgway, K.: Controlling the porosity of 316L stainless steel parts manufactured via the powder bed fusion process. Rapid Prototyp. J. 25, 162–175 (2019). https://doi.org/10.1108/RPJ-11-2017-0226.
18. Wang, R.J., Wang, L.L., Zhao, L.H.: Density prediction model of selective laser sintering parts. Hunan Daxue Xuebao/Journal Hunan Univ. Nat. Sci. 32, 95–98 (2005).
19. Yakout, M., Elbestawi, M.A., Veldhuis, S.C.: Density and mechanical properties in selective laser melting of Invar 36 and stainless steel 316L. J. Mater. Process. Technol. 266, 397–420 (2019). https://doi.org/10.1016/j.jmatprotec.2018.11.006.
20. Abreu, M.G. d., Pallone, E.M.J.A., Ferreira, J.A., Campos, J. V., Sousa, R.V. d.: Evaluation of machine learning based models to predict the bulk density in the flash sintering process. Mater. Today Commun. 27, 1–5 (2021). https://doi.org/10.1016/j.mtcomm.2021.102220.
21. Liu, Q., Wu, H., Paul, M.J., He, P., Peng, Z., Gludovatz, B., Kruzic, J.J., Wang, C.H., Li, X.: Machine-learning assisted laser powder bed fusion process optimization for AlSi10Mg: New microstructure description indices and fracture mechanisms. Acta Mater. 201, 316–328 (2020). https://doi.org/10.1016/j.actamat.2020.10.010.
22. Kosicki, J.Z.: Generalised Additive Models and Random Forest Approach as effective methods for predictive species density and functional species richness. Environ. Ecol. Stat. 27, 273–292 (2020). https://doi.org/10.1007/s10651-020-00445-5.

Comment 2. The amount of performed work in Results is limited. As in the first reviewer's response, it is recommended to extend the analysis by utilizing the model to a different set of additively manufactured parts or in another way that the authors consider appropriate.

Response: Thank you for the comment. We agree that the large amount of data enhances the prediction level of ML models. However, very limited experimental work reported related to process parameter variation for PBF-AM of SS316L. Thus, this work utilized the available experimental data of PBF-AM of SS316L and developed a density prediction model. Though the developed model is based on small dataset, it provides an understanding on density prediction which in turn helps in executing further analysis for large number of data. However, this aspect of availability of limited data has been included in the limitation of the conducted research. In near future, this model also can help to carry forward the similar studies with large data set and also for the other mechanical properties prediction.

Reviewer 2 Report

Thank you for your comment and supplementation. However, the label size of each figure is different.

-          Some legends are bold, and some are not (Figure. 7, 9, 10, 12).

-          The font form of the figure legend is also different (Figure. 9, 10, 14). Please unify the form.

It would be better to publish this manuscript after revising.

Author Response

Dear reviewer, 

We would like to thank you for your review. Below you can find our answers to your comments: 

     Comment 1. Some legends are bold, and some are not (Figure. 7, 9, 10, 12)

Response: Thank you for the comments.  In the updated manuscript, all the mentioned figures are updated with the uniform legends.

Comment 2. The font form of the figure legend is also different (Figure. 9, 10, 14). Please unify the form.

Response: Thank you for the comment. Figure no. 9, 10 and 14 legends have been formatted with the same font (Arial) with the same size (12) of legend. Thank you.

Comment 3. It would be better to publish this manuscript after revising.

Response: Thank you for the recommendation for publication. As per the reviewer’s comment, we have revised the manuscript.

Reviewer 3 Report

1. In Figure 6, the format of drawing notes (A) and (B) is inconsistent. It is suggested to select the same style for the horizontal and vertical axis of Figure 14;

2. Among the four machine learning model methods, the number of test sets used is different. There are three or four test sets, and the number of test sets is small. The fitting of linear regression equation is of little significance. At the same time, it is impossible to use a unified standard to evaluate the quality of the model;

3. All the pictures in the paper are not vector pictures, and the definition is not high. Try to modify them to high-definition and undistorted vector pictures;

4. In the paper, the R2 value of SVM, LR predicted density, RMSE value of KNN, etc. write decimal point as comma. For example 0,923;

5. How to calculate the actual percentage of density in the paper? Explain why the actual density percentage in SVM is lower than the predicted density percentage;

6. The writing level of the thesis needs to be improved. Many sentences use inappropriate words and are difficult to understand.

For example, in Section 3.3“hence few points are allowed to misclassify for obtaining a smooth hyperplane, as shown in Figure 6(B).”，The use of “few points”.

7. In this paper, SVM is the best method to predict the density, but the kernel function used by SVM and the optimization of cost value C and kernel parameter g are not described in detail, and the prediction results are not rigorous.

Author Response

Dear Reviewer,

We would like to thank you for your review. Below you can find our answers to your comments:

Comment 1. In Figure 6, the format of drawing notes (A) and (B) is inconsistent. It is suggested to select the same style for the horizontal and vertical axis of Figure 14;

Response: Thank you for the comment. In the update manuscript figures 6 and 14 are formatted with same style and removed the inconsistency.  

Comment 2. Among the four machine learning model methods, the number of test sets used is different. There are three or four test sets, and the number of test sets is small. The fitting of linear regression equation is of little significance. At the same time, it is impossible to use a unified standard to evaluate the quality of the model;

Response: Thank you for the comment. We agree that the large amount of data enhance the prediction level of ML models. However, very limited experimental work reported related to process parameter variation for PBF-AM of SS316L. Thus, this work utilized the available experimental data of PBF-AM of SS316L and developed a density prediction model.

Though the developed model is based on small dataset, it provides an understanding on density prediction which in turn helps in executing further analysis for large number of data. However, this aspect of availability of limited data has been included in the limitation of the conducted research. In near future, this model also can help to carry forward the similar studies with large data set and also for the other mechanical properties prediction. Also, due to the limited amount of data the training and testing data is divided into 70% for testing and 30% for training.

Comment 3. All the pictures in the paper are not vector pictures, and the definition is not high. Try to modify them to high-definition and undistorted vector pictures;

Response: Thank you for the comment.  We agree that the pictures in the article are not vector pictures. However, we ensured to include quality pictures that will provide clarity about the concept being discussed.

Comment 4. In the paper, the R2 value of SVM, LR predicted density, RMSE value of KNN, etc. write decimal point as comma. For example 0,923;

Response: Thank you for the comment, in the update manuscript all the numerical values are updated with the comma.

Comment 5. How to calculate the actual percentage of density in the paper? Explain why the actual density percentage in SVM is lower than the predicted density percentage;

Response: Thank you for the comment. The actual percentage of density mentioned in the paper is related to the experimental measurement of density with the ‘Archimedes principle’ as the ML models are able to predict the density well near to the actual density measured in experimental work. However, there is an MAE of 1.31 in prediction for SVM model. Hence the actual density percentage in SVM is lower than the predicted density percentage.

Comment 6. The writing level of the thesis needs to be improved. Many sentences use inappropriate words and are difficult to understand. For example, in Section 3.3“hence few points are allowed to misclassify for obtaining a smooth hyperplane, as shown in Figure 6(B).”，The use of “few points”.

Response: Thank you for the comment. The writing level of manuscript has been improved and inappropriate words are replaced.

Comment 7. In this paper, SVM is the best method to predict the density, but the kernel function used by SVM and the optimization of cost value C and kernel parameter g are not described in detail, and the prediction results are not rigorous.

Response: Thank you for the comment. More details about SVM has been included in the revised manuscript with following text.  

“The SVM model was validated with RMSE value with different cost functions as shown in Figure 11. The cost value ensures the smoothness of the hyperplane by selecting the different cost values. This graph shows the optimum value of the cost function for minimum RMSE value. The minimum RMSE value of 2,096 was obtained at the cost value (C) equal to one. The SVM models are easy to formulate to improve the model's accuracy.”