New NOMAD Publication
Aldo Glielmo, Peter Sollich and Alessandro De Vita published a novel scheme to accurately predict atomic forces using Gaussian process (GP) regression: “Accurate interatomic force fields via machine learning with covariant kernels” in Phys. Rev. B 95, 214302 (2017) .
The work introduces a number of new tools and ideas, including (i) the use of covariant matrix valued kernels to machine-predict very accurate interatomic forces directly as vector quantities; (ii) a natural definition and derivation of a general “two-body” force field for any target system, generated from a linear base kernel by analytic integration over the SO(3) group; (iii) a discrete summation technique to effectively restrict the covariance of the predicted forces to any desired finite symmetry group. This enables highly resolving symmetry efficient kernels e.g., based on the squared exponential form, or any other form for which analytic integration over the full SO(3) is not viable, to investigate crystalline solids.
The target accuracy of these potentials is high. Predicted forces were on average within ~0.1eV/A from the reference QM forces for a variety of systems tested (cf. figure below). This consistently improves on the existing “embedded atom method” force fields in metallic systems such as bulk and defective Fe and Ni.