Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

O: Fachverband Oberflächenphysik

O 67: Frontiers of Electronic Structure Theory: Focus on Artificial Intelligence Applied to Real Materials 3

O 67.6: Vortrag

Donnerstag, 8. September 2022, 11:45–12:00, S054

Equivariant N-center representations for machine learning molecular Hamiltonians — •Jigyasa Nigam, Michael Willatt, and Michele Ceriotti — Laboratory of Computational Science and Modeling, Institute of Materials, Ecole Polytechnique Federale de Lausanne, 1015 Lausanne, Switzerland

Most of the widely used machine learning schemes that have been successful in predicting chemical and material properties rely on concise, symmetry-adapted descriptions of the underlying atomic structure. A class of these structural descriptions is built on hierarchical correlations of atom-centered densities(ACDC)[1]. These are subsequently used to model corresponding atomic properties or atomic contributions to a global observable. However, many quantum mechanical quantities, such as the effective single-particle Hamiltonian written on an atomic-orbital basis, are associated with multiple atom-centers. This effectively renders ACDCs inadequate to describe the additional degrees of freedom of such multicenter properties. We recently proposed an N-centered representation[2] that extends the ACDC framework to the case of targets that are simultaneously indexed by N atoms. I will demonstrate how devising a family of N-center representations opens avenues for new classes of machine learning models that are fully equivariant and describe their role in assisting electronic structure calculations.

[1] J. Nigam, S. Pozdnyakov, M. Ceriotti, JCP 153,121101, 2020

[2] J. Nigam, M. Willatt, M. Ceriotti, JCP 156, 014115, 2022