DeepGraphene

Paper: Deep Learning Bandgaps of Topologically Doped Graphene Yuan Dong, Chuhan Wu, Chi Zhang, Yingda Liu, Jian Lin, JianLin Cheng
Url: https://arxiv.org/abs/1809.10860 < Script Form >
Author: Chuhan Wu

Information:

Author	Chuhan Wu	Dong Yuan	JianLin cheng	Jian Lin
E-mail	[email protected]	[email protected]	[email protected]	[email protected]

Introduction:

• This Repo contain the source code of Paper Deep Learning Bandgaps of topologically Doped Graphene , it contains all algorithms which we use to predict graphene supercells' bandgap values (Graphene-SVR, VCN, RCN, CCN). Meanwhile it contains the latest data of graphene supercell ( 4by4: 13018, 5by5: 79647, 6by6: 6382).
• DeepGraphene is an interdiscipline research that implemented Machine Learning methods toward the bandgap values prediction problem. It described different type of Graphene supercell structure into 2-D matrix, them input these data into Deep Networks or SVR algorithm to extract their spatial and hidden features. Therefore we can predict graphene supercells data band-gap values based on its 2-D structure matrix.

Brief workflow :

How to describe Graphene supercell structure:

The workflow chart (VCN)

Requirement:

* Tensorflow (1.11.0)
* Keras (2.2.4)
* Matlab (2017a version)

Index :

• Data set : (This folder contain the data of Graphene supercells)
  • Data
  • data_script
  • Original_data
  • Processed_Dataset
• Predict_h5file : (This folder contain all Deep neural netowkr models we have trained [except Graphene_SVR])
  • h5_file.zip
• Script : (This folder contain all scripts we use)
  • Predict
    • DeepGraphene (This folder contains Deep Neural Network algorithms' script: VCN, RCN, CCN)
    • Graphene_SVR (This folder contains traditional machine learning algorithm' script: Graphene_SVR)

Usage:

• Clone the whole repo into your local address.

• If you want to train DeepGraphene Neural netowrk. Go into Script folder , click Master.py . The line 23 to line 29 You can set which algorithm you want to choice: VCN, RCN or CCN, if you select one, please annotate other algorithms.
  • Meanwhile you can set the epoch you want to train and whether you need to use Transfer Learning toward the problem, transfer learning is good at single size training problem, if you want to do that. Set VCN as an example, please go to line 45 to anti-annotate this line and annotate line 46 too.
  • This Script is Training all graphene data together (4by4, 5by5, 6by6 data together) and the testing data are randomly selected from these data, each size have 1000 test data. The performance of predicting test data will show in the console window after you have trained this model.
  • Once your DeepGraphene algorithm's training process is finished, this model will preserve in this folder and named as total_TF_'algorithm you choice'.h5 (with transfer learning) or total_Non-TF_'algorithm you choice'.h5 (without transfer learning) automatically.
    

• If you want to train Graphene_SVR, Go into Script folder, Use Matlab2017a to open the script SVM_Regression. From line 8 to line 18 you can select the size of Graphene structure you want to train, please do the annotate and anti-anotate process toward this command line. After that,run it,them you can get the model you want, the performance will show in the windows after you have trained the model.

Result:

Setting 4by4 and 5by5 data as an example, the training process toward testing data is shown as below training single Graphene size data's performance