---
titles: KSE 2020 Answerings
tags: Paper_KSE2020
---
[TOC]
## Reviewer 1
1. References should be given for equations whitch are not derived or proved by the authors.
2. Figures and tables should be at the top or the bottom of the page. Done fix the tables and figures.
3. Each part of Fig.3 should be labeled using a), b), c) and so on. Also, each part must have a separate explanation in the caption of the figure.
4. The 'future work' selection should be added after conclusion.
5. Citations in reference section should follow IEEE format if it is not specified or predetermined.
6. References in reference section should be ordered based on their apperance in the content of the paper.
## Reviewer 2
1. How did the authors visualize local information and propagate interaction between atoms?
- Answer: For visualizing the local information and propagate interaction between atoms, we can look into the value of the filters (or egdes in the molecular graph). The values is represent for the interaction between two atoms in pair, so, by looking the value of all filters we can see how the atom interaction with the others.
2. Did the authors observe the training time between the proposed model and DFT?
- Answer: In general, the time gap between Deeplearning model and the DFT methods is huge. For Deeplearning models, we just need a little time (from 10 - 100 hours) to train models in the datasets and we can you this for predict very fast. But DFT methods take a huge time to calculate (days or weeks per molecule). So this is one of the motivation of applying Machine Learning, Deep learning models in predict molecular properties. In particular, our model is train seperate for each propertie and it take about 10 - 16 hours training with 1 K40m GPU (a little time and resource).
3. I think the authors should include the comparision in terms of MSE and LogMSE with DFT on the two datasets.
- Answer: As we show above, in this project, the Deeplearning models were trained in the datasets calculated by DFT methods. This is insprised that we can create the model that required not much data for training and even better so we can use this for predict. Our research continutes this approach so we compared our model with another state of the art Deeplearning models.

* **Comments**
* The authors should remove names in the current version. Yes I changed Linet to NAGCN (Node Aware Graph Convolution Neural Network).
* The motivation for adding the Readout layer is not clear enough. I suggest explaining why the authors used this layer.