Classifying Degraded Three-Dimensionally Printed Polylactic Acid Specimens Using Artificial Neural Networks based on Fourier Transform Infrared Spectroscopy

- In the introduction, the author mentioned that FTIR is one of the   best methods for understanding the degradation of PLA. The author should list   more references and give a short discussion about different methods and what   the advantage is when using FTIP method.

FTIR spectroscopy is a non-destructive and quick technique to access   molecular-level change of materials so it is easy to generate big data. In   order to observe changes in the molecular level, there are several other   technologies such as Raman spectroscopy, mass spectrometry, nuclear magnetic   resonance spectroscopy (NMR), and X-ray photoelectron spectroscopy[11].

- Section 2.1 Does the filling pattern affect degradation? If yes, did   author try other patterns?

I could say ‘yes’ in aspect of the   mechanical property. I do not think so for FTIR spectroscopy.

- Section 2.1 How to choose the measurement locations on the PLA   specimens?

We randomly selected them under the specific portion.

 I substituted three measurement   locations to three random measurement locations.

- Section 2.2, author mentioned that there are two control factors,   one is time and the other is temperature. However, the time for all sample is   set to 24h. So here should be only one control factor.

To solve the misunderstanding, I added one sentence   after that.

“This study focuses on one control factor of the   temperature.”

The paper deals with classifying degraded   three-dimensionally printed  polylactic acid specimens using Fourier   Transform  Infrared Spectroscopy and Artificial Neural Networks,   reporting interesting results. In the introduction of the work, the authors   state “Additive manufacturing or three-dimensional (3D) printing is widely   used in several industrial areas such as automotive, aerospace, mechanical,   medicine, biological systems, and food supply  chains [1]. Based on 3D   CAD data, additive manufacturing can be used to fabricate complex geometries   economically with a wide variety of materials. Polymers are the most commonly   used  materials in the 3D printing industry owing to their diversity and   ease of adaptation to different 3D  printing processes [2]…”.   Accordingly, I also suggest to BRIEFLY introduce some progresses in the field   of Additive manufacturing/three-dimensional (3D) printing for different   applications such as micro/nanocomposite scaffolds for tissue engineering   (i.e., “Three-dimensional printed bone scaffolds: The role of   nano/micro-hydroxyapatite particles on the adhesion and differentiation of   human mesenchymal stem cells”. Proceedings of the Institution of Mechanical   Engineers, Part H: Journal of Engineering in Medicine 2017, 231(6), 555-564   …) as well as the possibility to use recycled polymers to fabricate   3D printed structures (i.e., “A comparison between mechanical properties   of specimens 3D printed with virgin and recycled PLA” Procedia CIRP, Volume   79, 2019, Pages 143-146 …).

Copinet et al. [5] tried to estimate the   biodegradation of a co-extruded starch/poly (lactic acid) polymeric material.   Lanzotti et al. [6] studied the mechanical properties of virgin and recycled PLA. Currently, fused filament fabrication   (FFF) using PLA is commonly used by the public [7].

The list of references should be improved.

Lines   3 – 5 (article title): the title needs to be clarified, as current one   suggests classifying with FTIR and ANN, for example, the form:  Classifying degraded   three-dimensionally printed polylactic acid specimens using Artificial Neural   Networks on the basis of/based on Fourier Transform Infrared Spectroscopy can   be used.

“Classifying degraded three-dimensionally   printed polylactic acid specimens using Artificial Neural Networks based on   Fourier Transform Infrared Spectroscopy”

Lines   17-18: the sentence needs to be rewritten

“A   classification methodology using artificial neural networks (ANNs) based on   Fourier transform infrared (FTIR) and was developed.”

Line   55: the sentence needs to be corrected

“Zhang [14] reported that the thermal   degradation of 3D-printed ABS and PLA could be classified using ANNs based on   FTIR.”

Line   57: the sentence needs to be corrected

“The present study focuses on the thermal   degradation of 3D-printed PLA and classifies it by using ANNs based on FTIR.”

Line 148: the sentence needs to be corrected

“By using experimental data, this study   demonstrated herein that thermal degradation of 3D-printed PLA could be   classified using ANNs based on FTIR.”

Generally,   data used for ANN teaching especially when some signal is analysed, need to   be normalized, which means that the Author should conduct pre-analysis of   FTIR spectra covering baseline correction and normalising the spectra using   the greatest intensity or the area under the main peak. Not doing it may   result in a less reliable and less stable output. Of course, normalisation is   not mandatory when we have got binary data, however in that case   normalisation should be done. When the ranges of values for each input are   the same, tested model is more stable and modelling results are more   reliable. The Author should treat that comment as a clue for future work.

“Currently, this methodology has used FTIR   datasets without any preprocessing such as normalization. This study assumed that   ANNs could classify raw FTIR spectra. However, the preprocessing of the   datasets may give better performance to ANNs so that it could be further   work.”

Section 2.4: What was the way in which the   particular network architecture was chosen – has the range of ANNs was   tested, if yes, which algorithm has been applied for testing and on which   basis the best one was chosen. Has the one which was used proved to be the   best for that task? Was the only one architecture of MLP (presented in figure   4) used to classify all the data? Which learning algorithm was applied for   the chosen ANN (standard backpropagation or other)? The exact number of   neurons in all the layers should be given either in the text or in the   description of figure 4 (inputs number depending on the division of the   training set, the number of neurons in the hidden layers and in the output   one).

I need to reduce the number of   hyperparameters to solve this study. So I need to fix several hyperparameters   guided by J.Heaton’s book “ Introduction to Neural Networks for Java”. According   to the rules to generate the ANN model, the number of neurons varied with the   data size. There were five cases as shown in Figure 7.

Line 90: Artificial Neural Networks is the   name for the big family of networks, while Multilayer Perceptron is only one   of the members of that family, next to Radial Basis Function Networks,   Kohonen networks etc. The sentence needs to be rewritten.

 “Multi-layer   perception (MLP) as artificial neural networks (ANNs) was used to classify   the degree of thermal degradation of the PLA specimens used herein.”

Lines 98-99: Could you explain why as an   activation function rectifier linear unit was used, not the other one? Has   any other been tested? Please, develop the shortcut used.

“ In   this study, an activation function for the nodes was ReLU[15] which   is currently known as the most successful and widely-used activation function[16]. “

In   the description of figure 6 please add the symbols of degradation levels   (D01-D08).

I added the symbols for figure 6.

Line   128: I do not understand why did you use only four outputs as you need to   distinguish among eight degradation levels. Have networks been not able to   conduct such a classification? Why only D01, D02, D07 and D08 were chosen?   Were the FTIR spectra to similar in other groups? I do not see any probable   explanation of that problem, which seems to be crucial for designing reliable   and independent classifier.

The maximum number of classifications   was four under this study. The ANN models could only classify D01, D02, D07,   and D08, when the number of divisions in FTIR spectra ranged between four and   sixteen. More datasets and preprocessing of FTIR spectroscopy may increase   the number of classifications. It will be further work.

The   technique used for finding the most proper wavenumber ranges for the   classification is quite tricky. Scanning the spectra seems to be the correct   attitude, especially if you do not know characteristic peaks. However, have   you tried to use as an input only those ranges that cover the main peaks for   the neat PLA, for example, 750-760 cm^-1, 865 to 875 cm^-1 or   others?

This classification method was initially   designed and suggested for novices who are not familiar with FTIR spectroscopy.   If we want to use the small range (750-760 cm^-1 or 865 to 875 cm^-1 ),   it appears that we need to apply any data processing such as normalization.