The implementation of conditional GAN for the prediction of the transition state (TS) geometry based on cartesian coordinates of product and reactant. For more information, please refer here to our publication.
The updated version of the TS-GAN (predictions) that uses the up-to-date TensorFlow 2.12 and Atomistic Simulation Environment (ASE) to speed up the code is available in the updated branch.
- Python 3.8.16
- Tensorflow 2.2.0
- Numpy 1.20.0
Go to the working directory:
git clone https://github.com/ekraka/TS-GAN.git
To predict the TS guess structure, make sure the g_model.h5 file is in the same working directory as xyz files of reactant and product.
The g_model.h5 can be found in the test_cases folder depending on which reaction one is interested in.
python predict.py reactant.xyz product.xyz
Prediction script will generate two files: temp_ts.xyz and temp_mov.xyz. The first file shows the final guess structure, while the second file shows the movie on how the optimization took place.
To train the model on your own data, convert xyz files of reactants, transition states, and products into the Coulomb matrices (CMs) and store as numpy file. This can be done with the gen_data.py script which requires specific format for files. The reactnats, transition states and products should be kept in a single folder with the following name specification:
Transition state: filename.xyz
Reactant: filename_rev.xyz
Product: filename_for.xyz
To create the numpy data, run the following command from this directory:
python path_to_TS-GAN/gen_data.py
This will create a file called data.npy
To train through this data, create a prefered working directory and copy the data.npy file. Also, create an empty folder temp in this directory, then run:
python train.py
During the training process, the model will save weights of discriminator and generator as d_model.h5 and g_model.h5, respectively. While the random real and fake samples of the CMs are saved in the temp folder.
To calculate the root-mean-square deviation (RMSD) use:
python align3D.py ts.xyz temp_ts.xyz
M.Z. Makoś, N. Verma, E.C. Larson, M. Freindorf, and E. Kraka; Generative Adversarial Networks for Transition State Geometry Prediction; J. Chem. Phys. 155, (2021), Vol.155, Issue 2; doi.org/10.1063/5.0055094
@article{TSGAN,
doi = {10.1063/5.0055094},
year = {2021},
publisher = {{AIP} Publishing},
volume = {155},
number = {2},
pages = {024116},
author = {Ma{\l}gorzata Z. Mako{\'{s}} and Niraj Verma and Eric C. Larson and Marek Freindorf and Elfi Kraka},
title = {Generative adversarial networks for transition state geometry prediction},
journal = {J Chem Phys}
}