Novel Auxiliary Options¶
Type embedding¶
Instead of training embedding net for each atom pair (regard as G_ij, and turns out to be N^2 networks), we now share a public embedding net (regard as G) and present each atom with a special vector, named as type embedding (v_i). So, our algorithm for generating a description change from G_ij(s_ij) to G(s_ij, v_i, v_j).
- We obtain the type embedding by a small embedding net, projecting atom type to embedding vector.
- As for the fitting net, we fix the type embedding and replace individual fitting net with shared fitting net. (while adding type embedding information to its input)
Training hyper-parameter¶
descriptor:”type” : “se_a_ebd” # for applying share embedding algorithm”type_filter” : list # network architecture of the small embedding net, which output type embedding”type_one_side” : bool # when generating descriptor, whether use the centric atom type embedding (true: G(s_ij, v_i, v_j), false: G(s_ij, v_j))
fitting_net:”share_fitting” : bool # if applying share fitting net, set true