Classifying Thermal Degradation of Polylactic Acid by Using Machine Learning Algorithms Trained on Fourier Transform Infrared Spectroscopy Data
Round 1
Reviewer 1 Report
The manuscript presents a new research of the author on the subject of experimentally analyzing the thermal degradation of PLA based on FTIR data; this time two spectometers are used and four machine learning algorithms provided by Microsoft Azure Machine Learning Studio (MLS) are tested for prediction.
The research is experimental and the practical implication of the results is missing. The author should clearly state what the practical purpose of this research is. It will be very useful to give examples of situations in which 3D printed PLA objects are subjected to high temperature (160C) for long periods of time (24h). Usually, when the functional specifications include such extreme conditions, other types of materials are used. Moreover, in my opinion, analyzing the effect of thermal degradation on 3D prints mechanical properties and dimension and form accuracy is more important for “manufacturing robust products by using 3D printing technologies” as the author mentioned in abstract, than just an assessment of the degree of thermal degradation. Therefore, it is confusing why type IV tensile specimens were needed for this research.
In section 3 results are presented, however there is no discussion in the context of data from other research. For instance, the author could have discussed the results of the experiments in the context of the previous paper from 2018 which analyzed ANN. Why the author considered necessary to test new algorithms and, according to section 4, to continue the work on the same line/approach with two other training strategies? Somehow this leaves the impression that the research included in this manuscript is only partial disseminated. Therefore, my suggestion is to add also the information from testing other training strategies in order to offer a complete, integrative perspective over the subject, and then resubmit the paper. More data are needed to enhance the value of the paper.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The paper presents the prediction accuracy of the thermal decomposition of PLA based on FTIR data by different machine learning algorithms. The paper is well written and summarizes the results very concisely. I am not an expert in machine learning, so I cannot say much about these aspects. In general, I think the topic is interesting for the scientific community. As FTIR is one of the most frequently used characterization methods, the approach is likely to be relevant for experts in this field and could possibly be applied to other measurement methods, too. After some minor adjustments and additions, I would recommend publication in Applied Sciences.
ll. 25-32: The text has not been deleted from the template.
l. 44: The classification and the following explanation need references.
Table 1: The abbreviations should appear after the respective term. It is unclear whether “mechanical test” does always refer to impact strength, hardness and elastic modulus. The use of abbreviations is not consistent in the table.
The comparison with other machine learning applications in the field of material characterization could be broadened.
ll. 92-97: Relative humidity should be stated.
l. 110: Some experimental information is missing: Have the spectra been smoothed or corrected for atmospheric noise by software? What kind of measurement method has been used, ATR (if so, what crystal material?) or transmission?
ll. 117-118: Normally desktop devices average over the repeated measurements. It is unclear why the results were picked randomly instead of averaging.
Figure 3: The spectra are way too small, it’s literally impossible to read the legend. A slight offset of the data could help distinguish the graphs. Which specific signal changes in the spectra could be expected due to thermal degradation should be discussed briefly and at least one reference should be cited.
Figure 7. If appropriate the resulting degradation should be briefly compared to those studies cited in the introduction. Colorimetry is common method to evaluate degradation, it should at least be mentioned because the micrographs are evaluated without objective measurements. Since 3D printed parts are often colored, the role of quick and inexpensive chemical evaluation such as FTIR data is especially important.
The applicability for other characterization methods could be briefly highlighted in the conclusion.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
In this report the ageing, or thermal degradation, studies are conducted on PLA prints obtained by fused filament fabrication method. FTIR was used to monitor the accelerated degradation process where the raw FTIR data was fed into four machine learning algorithms.
Analysis of the accuracies of the prediction models helped in determining that multi-class logistic regression and multi-class neural performed better than other two methods.
The most interesting data was extracted from the confusion matrices for the multi-class logic regression and multi-class neural network methods for the training FTIR datasets used in these study.
Overall, it was shown that both algorithms show good predictive agreements for all FTIR datasets except for one, i.e. where the FTIR spectra were collected in time-controlled storage test. This was the second training data set.
This is an interesting study, clear and concise. It can probably be adopted for FTIR degradation prediction studies of many other polymeric materials. In my opinion it can be published after revising.
Below are other comments that the author should consider addressing in a revision.
It is not clear why the PLA was 3D printed. Dog-bone shaped samples are typically used in mechanical testing. Does the author have the capacity of running such tests? FTIR could have been performed on samples in any other form. Was the IR study done in the ATR-IR mode? A printing texture can be seen in the treated samples. Does that not affect the spectra collection? This is not clear in the manuscript.
The IR spectra in Figure 3 are not clear to see any differences resulting from the ageing. I would suggest to show only representative scans (3-5) including initial and final ones, and at better resolution. That way the reader could see if there are any differences in IR signals.
Is this degradation time/temperature sufficient to observe any degradation at all rather than some discoloration? I think the largest differences should be emphasized in the respective spectra.
Is it possible to use only partial spectral ranges for the learning process and simulations? Just like the author did in his previous publication (Appl. Sci. 2019, 9, 2772). Would a better agreement be expected when more relevant peaks/data are used for machine learning?
Are there any references on using the applied algorithms in any other machine learning processes?
The first paragraph of the introduction seems to be part of a template.
Line 226-227: there is “models trained on the 3rd dataset (Figure 9)”, but Figure 9 depicts 2nd dataset.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The author did not consider the most important remarks/comments from my review.
Only two minor modifications were brought to the manuscript based on my observations. One of them refers to the use of tensile IV specimens (What is the thickness of the sample? Why not performing experiments of PLA filament?), which indeed can be found in Choi et al. study, but because they have measured the mechanical properties. Therefore, the added text is not an argument/answer to my comment. The other addition/modification briefly explains the difference between this study and the author’s previous studies on the same topic. However, the information in section 3 is still not related to other research as I suggested in my comments. In my opinion, in this manuscript the results are presented, but not discussed/interpreted in the context of other works or in the context of the 3D printing process. No references to other studies can be found in section 3. How the obtained results are influenced by the fact that the specimens are 3D-printed?
The author did not explain the practical implications of the research, in my opinion very important aspect considering the aims and scope of the Applied Sciences journal. Adding a paragraph on how these results can be used in real case scenario would have been beneficial. What practical information this research conveys to a 3DP user/designer? Especially when the conclusions mention that other training strategies and normalization methods can also be suitable to achieve the same purpose: “Moreover, other training strategies such as hyperparameter tuning, other normalization methods, and applying partial spectral ranges could similarly enhance the predictive capacity of models and are the subject of future studies.”
The experiments were performed on one PLA material, not on various PLA materials. Plural is not correct here, misleading the reader who might expect results from testing on two or more PLA materials from different producers, colors etc.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
Figure 3 in its present form is unreadable and should be removed.
The comments were addressed in part.
Some questions were left unanswered.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 3
Reviewer 1 Report
-
Author Response
Please see the attachment.
Author Response File: Author Response.docx
