Fig. 1: Sketch of neural network architecture.
From: Ab initio calculation of real solids via neural network ansatz
The electron coordinates ri are passed to two channels. In the first one, they build the periodic distance features d(r) using the periodic metric matrix M and the lattice vectors a, and then d(r) features are fed into two molecular neural networks, that represent separately the real and the imaginary part of the wavefunction. In the second channel, ri constructs the plane-wave phase factors on a selected set of crystal momentum vectors. The total wavefunctions of solids are constructed by the two channels following the expression of Eq. (1).
