Parts | Days | Selection | Search | Updates | Downloads | Help

MM: Fachverband Metall- und Materialphysik

MM 34: Data Driven Materials Science: Interatomic Potentials / Reduced Dimensions

MM 34.5: Talk

Thursday, September 8, 2022, 16:45–17:00, H45

Kernel Charge Equilibration: Learning Charge Distributions in Materials and Molecules — •Martin Vondrak, Nikhil Bapat, Hendrik H. Heenen, Johannes T. Margraf, and Karsten Reuter — Fritz-Haber-Institut, Berlin, Germany

Machine learning (ML) techniques have recently been shown to bridge the gap between accurate first-principles methods and computationally cheap empirical potentials. This is achieved by learning a systematic relationship between the structure of molecules and their physical properties. However, the modern ML models typically represent chemical systems in terms of local atomic environments. This inevitably leads to the neglect of long-range interactions (most prominently electrostatics) and non-local phenomena (e.g. charge transfer), which can lead to significant errors in the description of polar molecules and materials (particularly in non-isotropic environments). To overcome these issues, we recently proposed a ML framework for predicting charge distributions in molecules termed Kernel Charge Equilibration (kQEq). Here, atomic charges are derived from a physical model using environment-dependent atomic electronegativities. These models can be trained to reproduce electrostatic properties (e.g. dipole moments) of reference systems, computed from first principles. The impact of different fitting targets on predicted charge distributions is compared. Furthermore, strategies for fitting to energies are discussed, including combination of Gaussian Approximation Potential (GAP) with kQEq.