Berlin 2018 – wissenschaftliches Programm

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

CPP: Fachverband Chemische Physik und Polymerphysik

CPP 49: Soft Matter Physics: Emerging Topics, New Instruments and Methods

CPP 49.5: Vortrag

Mittwoch, 14. März 2018, 16:00–16:15, C 230

Symmetry-Adapted Machine-Learning for Tensorial Properties of Atomistic Systems — •Andrea Grisafi, David Mark Wilkins, and Michele Ceriotti — Laboratory of Computational Science and Modeling, EPFL, Lausanne, Switzerland

In the last few years, machine-learning methods have played a prominent role in providing an accurate description of atomic-scale properties, bypassing computationally demanding quantum chemistry calculations.

A great deal of effort has been applied to learning the ab initio ground state energy of materials and molecules, giving accurate potential energy surfaces. However, a complete description of a condensed phase system requires the calculation of properties that are tensorial in nature, e.g., multipole moments. Unlike scalar quantities, these properties transform in prescribed ways when the reference frame is rotated, presenting a great challenge when designing a machine-learning algorithm.

I discuss an extension of a traditional machine-learning method which can be used to predict tensorial properties of arbitrary rank for systems of arbitrary complexity, automatically encoding the rotational symmetry in three dimension. The novel algorithm has the potential to cover a broad range of applications within material science and condensed matter physics. Applications are shown for the prediction of the electrical response series of water oligomers, from the single molecule to the condensed-phase. Preliminary results for the prediction of the ab initio charge density of organic molecules will be also presented.