Fig. 1: Schematic representation of the moment tensor potential (MTP) and the Gaussian moment neural network (GM-NN) approaches.
From: Performance of two complementary machine-learned potentials in modelling chemically complex systems
a Both methods rely on the assumption of locality of interatomic interactions but implicitly include interactions beyond the cutoff radius Rcut. b As an intermediate step in constructing invariant features, both methods obtain a common representation equivariant to rotations. c Invariant features are obtained by computing tensor contractions and eliminating linear dependencies specific to the considered approach. d Utilizing linear regression or neural networks, which take invariant features as an input, mapping from a structure to the respective energy is modelled.
