Construct a distribution-to-distribution (DTD) model for a reactive atom-diatom collision system. Here, we consider a grid-based DTD model for data from Set1 in reference arXiv:2005.14463.

Requirements

python3, TensorFlow and matplotlib

Data

The data folder contains files for 9 sets of reactant and product state distributions from Set1 that were obtained by quasi-classical trajectory simulations. The corresponding sets of temperatures , can be read from the file tinput.dat.

Prepare input and reference values for training a neural network

Run the code generate_input_and_reference.py located in the folder data_preprocessing to generate the files containing the input and reference values as well as a PDF with the corresponding plots, based on the 9 data sets from the folder data.

Train a neural network

The folder training already contains a file with input and reference values that were generated using the code generate_input_and_reference.py considering the complete Set1, with and .

To train a neural network (NN), edit the code training.py in the folder training to specify the NN architecture and hyperparameters. Then run the code training.py by mentioning number of training data sets, validation data sets, seed, number of training epochs and batch size:

python training.py 3000 600 11 2000 500

Get the optimized neural network parameters

After finishing with the training, edit the code print_coeff.py in the folder training to load the NN model resulting in the lowest validation loss. Run the code with the same parameters as training.py to obtain the values of the corresponding optimized NN parameters in separate files:

python print_coeff.py 3000 600 11 2000 500

Construct the predicted product state distributions

Edit the code evaluation.py in the folder evaluation to specify whether and what accuracy measures (RMSD, ) should be calculated. Then run the code to obtain a file containing the desired accuracy measures, as well as a PDF with the corresponding plots for data sets from the folder data.

Cite as Julian Arnold, Debasish Koner, Silvan Kaeser, Narendra Singh, Raymond J. Bemish, and Markus Meuwly, arXiv:2005.14463 [physics.chem-ph]