Graph Database

We have included here an addition to the normal water cluster database which attaches a graph representing each water cluster to the relevant cartesian coordinates.

The paper associated with the generation and analysis of these graphs is:

Jenna A. Bilbrey, Joseph P. Heindel, Malachi Schram, Pradipta Bandyopadhyay, Sotiris S. Xantheas, and Sutanay Choudhury “A Look Inside the Black Box: Interpretation of a Continuous Filter Convolutional Neural Network (CF-CNN) for the Potential Energy Surface of Water Clusters using Graph-Theoretical Descriptors“, Journal of Chemical Physics 153, 024302 (2020).

These graphs are stored as networkx objects.

The final optimized weights for the neural network from the paper are available to download here. This potential can only be run by using SchNetPack. The settings we used in training can are in the same zip as the weights  in json format.

The complete database, with the graphs, can be downloaded from this link (13.6 GB). Additionally, a version of the database which only contains the projected graphs, i.e. graphs representing the oxygen framework, can be downloaded from here (800 MB). After unzipping the file, you will find multiple .zip files which are compressed text files containing all of the structures and graphs for that size of water cluster. You can decompress the file and then load a file with the name “example_graph.txt” into networkx objects using the following python code:

import networkx as nx
import json
f = open("example_graph.txt", "r")
s=f.read()
f.close()
s=s.split('---')[:-1]
G=nx.json_graph.adjacency_graph(json.loads(s[0]))