mileszhao / pgcgm

Source code for generating materials with 20 space groups using PGCGM

PGCGM:

Physics Guided Generative Adversarial Networks for Generations of Crystal Materials with High Symmetry Constraints

Created by Yong Zhao

Introduction

This repository contains the implementation of generation code of PGCGM. If you want to try the code quickly, you run it on Google Colab. The colab link is here PGCGM Colab. After click the link, first download the datamodel.zip file and upload to your colab drive. And then Runtime/runall menu.

The following installation guide is for running our model on your local machine.

Running environment set up

we recommend that you build a virtural environment running the code. Below are steps that can install the dependent packages. The installation has been tested under Ubuntu Linux with Nvidia 2080Ti and Nvidia 3090 GPUs.

Step1: Create conda virtual environment and activate it

conda create -n pgcgm python=3.7
conda activate pgcgm

use 'which' command to check folder of pip3 is located within the actual env pgcgm folder. It should be somewhere like if you are using miniconda:

which pip3
~/miniconda3/envs/pgcgm/bin/pip3

or

$HOME/.conda/envs/pgcgm/bin/pip3

Step2: Then install following packages using:

pip3 install pymatgen==2022.0.6
pip3 install pickle5

Install pytorch from pytorch web based on your python & cuda version. For example, For miniconda

pip3 install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

If your machine has no GPU, use this:

pip3 install torch==1.7.1+cu110 torchvision==0.8.2+cpu torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

Step3: Once you install all those packages, you just run below commands to generate crystal structures:

git clone https://github.com/MilesZhao/PGCGM.git
cd PGCGM
./sh/gen.sh
ls ternary_final_cifs

Your generated cif files are located in the ternary_final_cifs folder.

pgcgm's Issues

Clarification about model and results in paper

Hello and thanks for making this available!

I have a few clarifying questions about what you have here vs. in the paper.

1. In your main table of results (Table 1, https://arxiv.org/pdf/2203.14352v2.pdf), you have a few different PGCGM versions (PGCGM, PGCGM+dist, PGCGM+dist+coord). Which of these configurations corresponds to the model you've uploaded here https://github.com/MilesZhao/PGCGM/blob/main/models/frac12/generator_weights.pth) ?
2. Can you clarify what filtering criteria you used to obtain the set of MP structures here (https://github.com/MilesZhao/PGCGM/blob/main/data/mpids.csv)?
3. Similarly, would it be possible for you to add OQMD and ICSD IDs for the other data you used?

Thanks!

Do you have a comparison to CDVAE?

I noticed you compared results to FTCP. Do you have comparison results for CDVAE? I understand this might be less feasible for the 500k generation based on the longer time it takes to generate CDVAE samples.

about "paras.pickle" file

what's the datas of paras.pickle?

Use git to download the 'datamodel.zip' in the colab tutorial

hi, please consider using git to download the file, just like this:
$ git clone https://github.com/MilesZhao/PGCGM.git
$ mv /content/PGCGM/datamodel.zip /content
This is a little bit easier.
Thanks for your response.

Number of final CIFs

Hello,

Thank you for sharing your work!

I just run your code by setting number of samples to 1000 and I obtained 38 CIF files under ternary_final_cifs folder. How should I interpret these results? Does that mean the CIF conversion rate is 3.8%?