AP-Net: An atomic-pairwise neural network for smooth and transferable interaction potentials

This repository contains code and data associated with the AP-Net paper.

Dependencies

This project uses Python 3 and requires up-to-date versions of the following Python packages:

• numpy
• scipy
• pandas
• tensorflow (2.1.0 or newer)

Data

Datasets of dimer interaction energies are stored as serialized pandas dataframes named dimers.pkl in separate subdirectories of datasets/. Each row of a dataframe is an individual dimer. The columns RA and RB are monomer atomic coordinates (in Angstrom) and the columns ZA and ZB are monomer atomic numbers. For all datasets, the SAPT0/jun-cc-pVDZ interaction energy is stored in the Total column, and the four components that sum to this energy are stored in additional columns: elst, exch, ind, and disp.

To use your own dataset with AP-Net, add an additional subdirectory to datasets with a dimers.pkl file. RA, RB, ZA, and ZB are required columns.

AP-Net requires atom-pair descriptors for both inference and training. Instead of repeatedly recalculating these descriptors (also called features) on the fly, they are cached in the datasets/ subdirectory after creation for future use.

Using a pre-trained model

The trained AP-Net model described in the paper can be used to perform inferences on any included or added datasets. This is demonstrated in the script infer_sapt.py, which predicts the interaction energy (total SAPT0 and components) for all dimers in a dataset. Run as:

>>> python infer_sapt.py --help

for more information.

Training a new model

A new AP-Net model can be trained on an arbitrary dataset of SAPT0 dimer interaction energies. This is demonstrated in the script train_sapt_component.py which trains a single model to predict a single SAPT0 component. Run as:

>>> python train_sapt_component.py --help

for more information.