Skip to content

The PiNet network

The PiNet network implements the network architecture described in our paper.1 The network architecture features the graph-convolution which recursively generates atomic properties from local environment.

One distinctive feature of PiNet is that the convolution operation is realized with pairwise functions whose form are determined by the pair, called pairwise interactions.

Parameters

Parameter Default Description
atom_types [1, 6, 7, 8] List of elements
rc 4.0 Cutoff radius
cutoff_type 'f1' One of 'f1' or 'f2'
basis_type 'polynomial' One of 'polynomial' or 'gaussian'
n_basis 4 Number of radial basis functions to generate
gamma 3.0 controls width of the Gaussian basis
center None replace centers the for Gaussian basis with a list
pp_nodes [16, 16] specifies the property-property hidden layers
pi_nodes [16, 16] specifies the property-interaction hidden layers
ii_nodes [16, 16] specifies the interaction-interaction hidden layers
out_nodes [16, 16] specifies the output hidden layers
out_units 1 the dimension of outputs
out_pool False min, max or sum, pool atomic outputs to give global predictions
act 'tanh' activation function to use
depth 4 number of graph-convolution layers to use

  1. Y. Shao, M. Hellström, P. D. Mitev, L. Knijff, and C. Zhang. PiNN: a python library for building atomic neural networks of molecules and materials. J. Chem. Inf. Model., 60:1184–1193, January 2020. doi:10.1021/acs.jcim.9b00994