Machine Learning Based Prediction of Nanoscale Ice Adhesion on Rough Surfaces
Round 1
Reviewer 1 Report
Overall merit could be improved with additional explanation.
We greatly appreciate the time and the insightful comments from the reviewer!
1. As the authors realize the nano scale roughness they consider 70 nm or so is insignificant compared to even the smoothest surface aerodynamically 2micron about 20000 nm (up to 10mic 100000 nm) . And this is not really roughness in an operational sense.
This study indeed focuses on smaller (nano) length scale than experiments, specifically the nanoscale intrinsic ice adhesion, rather than the macroscopic properties where the commonly observed surface roughness play a role. Surface roughness at nanoscale affects the local properties of ice, and thus the ice adhesion strength locally. Such local ice adhesion should impact macroscopic ice adhesion, the manner of which requires more investigation and will be the focus of further studies. We emphasize the concern of the reviewer in the introduction section on Page 2 of the revised manuscript (text highlighted in red).
2. I am not an expert in this particular field so forgive my ignorance but as far as I know water creates polar bonds with surface, the way each molecule is simulated by Lj and a corresponding epsilon one would imagine that higher the epsilon stronger the bond, which does not seem to be the case.
We fully agree the higher LJ energy depth value, epsilon, should lead to high ice adhesion, if all the ice molecules are detached away from the substrate. In experiments, fully dry surfaces (without any residual ice/water molecule) after de-icing are very unlikely and mostly can only observed on superhydrophobic surfaces. In this study, residual water molecules were also observed on the surfaces. As such, the ice adhesion strength observed not only depends on the water-substrate interaction, but the properties of the ice/water layer adjacent to the substrate. If the structure of this interfacial water layer is more ordered (meaning more stable hydrogen bonding), the monitored de-icing force will be higher. The selection of water model and its interaction with substrates followed previous studies (ref. 9 & 50). As the water model is a coarse-grained model for the sack of calculation efficiency, it does only catch the interaction strength with the substrate but not every detail such as hydrogen bonding orientation and so on. We provided more comments on this regard in the Method section in this revision on Page 3.
3. I realize that (Figure S1) somehow for higher epsilon more amorphous water ! this is also interesting since there is no discussion of temperature or phase change . Where does the energy go . Normally since temperature corresponds to molecular kinetic energy there should be some relationship between epsilon, temperature, phase change .. Needs more explanation at least for me.
There are amorphous ice molecules (or quasi-liquid water) at ice surface or at the most nearby water layers close to a substrate below water freezing temperature (ref. 8, 9, 27, 32 & 33). Water molecules at such location are different from bulk. The results in this study (Figure S1) indicated the strong interaction between ice and substrate can increase the amorphous ice molecule number, meaning the stronger attraction force to substrate can more effectively disrupt the ordered hydrogen bond network close to the substrate. Despite the amorphous ice structure, the temperature of the whole ice/water molecules are the same. Individual water molecules can join into order hydrogen bonding network or remain in amorphous structure depends on the interplay between interaction strength with the substrate and hydrogen bonding energy with other water molecules. We further our comments on this point in the revised result section on Page 6.
4. Naturally (connected with my 1st comment) the training examples for ML seams limited.
We agree more training data and examples could greatly enhance the accuracy of our ML study in this work. Although this study has exploited the limit of our computing resources, further accumulation of data and more training algorithm will be tried out in the future steps of this study. We emphasize this point in the result section of this revision on Page 11.
5. I am not sure how one would use this idea / calculation to down select / evaluate different coatings. Or may be I am missing something. More vivid description of the use of the concept would add value to the paper.
As discussed above, this study only focused on nanoscale roughness at intrinsic contacts between ice and substrates. The results are limited to intrinsic ice adhesion. This study is thus the first step towards the desired multiscale prediction of ice adhesion on different surfaces for evaluating and selecting coatings in experiments. We further commented the motivation of this study in the discussion on page 11 of the revised manuscript.
Reviewer 2 Report
1. Line 44. Sentence is incomplete. Please correct it.
We have revised the incomplete sentence on Line 44 in our manuscript.
2. The adhesion test evaluated by the authors seems to be in purely uniaxial tensile mode. Are most mechanical de-icing practices/methods in this mode of fracture? Would a shear mode be more relevant and accurate representation of actual de-icing mechanism? It would be good to enlighten the audience about this.
Tensile and shearing at atomistic scale are both relevant to ice adhesion. Macroscale de-icing from a surface combines nanoscale separation of atoms ice molecules from the substrate atoms, featuring tensile separation, and then shearing of the interface. This study focused on local atomistic ice adhesion at intrinsic contacts area between ice and the substrate, which is the first step on multiscale prediction of macroscale ice adhesion. Although this study has exploited the limit of our computation resources for accumulating nanoscale ice adhesion data, further effort is needed to cover all the possible de-icing mechanics including shearing in the future study. We further comment this point in the revised method section on page 4 (text highlighted in red).
3. In the simulation the authors have chosen aluminum as the substrate. However, it would be good to provide a background on the role of substrate properties (elastic, thermo-mechanical etc..) on the de-icing mechanism and their corresponding challenges. This would help provide more context to the theoretical work carried out in this research.
We assume the substrate is much harder than ice this study. As such, the substrate is set to be fully rigid in the modeling for the sake of simplicity. To cover the change or wear of the substrates is currently not included in this work, which is yet very important in the future study. We have noted this important point in the method section of the revision on page 4.
4. What is the different mechanism in which de-icing occurs on substrates both at the macroscopic and microscopic level? Kindly include some more background to the overall fracture / failure mechanisms in the introduction section.
We have highlighted the understanding of ice adhesion mechanism at different length scale, especially the apparent and intrinsic ice adhesion, in the introduction of the revised manuscript on page 2.
5. How does the presence of preexisting defect (e.g.; debris, organic inclusions etc.) influence the de-icing mechanism on substrates. For instance, in the current work authors point to the role of a thin lubricating water layer’s role in affecting adhesion strength. Similarly, most surfaces that are susceptible to icing would not have pristine surfaces and contamination could potentially interact with the surface roughness to influence de-icing behavior. Kindly elaborate on these aspects.
We acknowledge the question raised by the reviewer on the effects of preexisting imperfect contact between ice and substrate on ice adhesion. Surface conditions greatly affect ice adhesion in experiment. Many important factors, such as contamination, voids and so on, act as a type of the so-called nanoscale crack initiators (ref. 21) which will degrade the intrinsic interface properties between the ice and substrate. However, studying the effect of the factors in detail exceeds the capacity of atomistic model was investigated in this study, and it should be considered in the relevant further investigations. We include new comments on this regard in the revised conclusion on page 12.
6. In the model and simulation that the authors have carried out, what would be residual stress state and order of magnitude of the ice layer. Once again since there are different phases of water/ ice, their stress states would be expected to change on substrate material and conditions. For the benefit of the wider audience, kindly shed some light on these aspects
We have further commented in the results section (page 6) on the adhesion of ice on surface with varied interaction strength and their effects on the structure of the interfacial water layer. The results indicated that the stronger attraction between ice and substrate led to more obvious disrupted ice structure adjacent to the substrate, which further significantly affect ice adhesion.
Reviewer 3 Report
The manuscript will be very interest to the readers. The machine learning based prediction of nanoscale ice ashesion was used to prevent icing events. The manuscript is well organized and all data are good for the final conclusion.
So I recommend it accepted as in the present form.
We greatly appreciate the time and kind recommendation!
Reviewer 4 Report
The paper “Machine Learning based prediction of nanoscale ice adhesion on rough surfaces”, by Simen Ringdahl, Senbo Xiao, Jianying He and Zhiliang Zhang, presents encouraging results upon ice adhesion prediction at nanoscale order. The paper is very-well written, and the information is clearly presented.
The paper is recommended for publication with minor revisions:
1.In section 2, the authors present rigorous, details related to rough surfaces, but nothing is mentioned regarding flat surfaces. The flat surfaces appear starting with section 3, in the results. Please insert some mentions about flat surfaces in section 2.
We have further discussed the results regarding ice adhesion on flat surfaces in the method section (section 2) on Page 3 (text highlighted in red) in our revised manuscript.
2. Have you considered to use a well-defined roughness value, not only randomly?
We initially only considered random roughness that corresponds to surfaces found in nature. We also believe that well-defined roughness is interesting and highly relevant to man-made surfaces, especially those found in modern electrical devices. However, the study of well-defined roughness may not need the machine learning. We have further commented this important point in the revised result section on page 12.
3. The authors were focus on the hexagonal ice. Due to the fact that in nature we have also other morphologies in the same time, how it could be influenced the adhesion on rough surfaces, if we take into account 2-3 forms of the ice, not only one? It could be interesting in the future to have some simulations by using various forms of the ice.
We fully agree with the reviewer that adhesion of different ice form should be also addressed. Although the current study focuses on hexagonal ice, different ice forms should be considered in the future steps for formulating a multiscale prediction framework of ice adhesion by machine learning. We have added comments on this regard in the results section on page 12.
Round 2
Reviewer 1 Report
Already received by the authors for further research ..
