CCMP Seminar Series - Bruno von Bruening
When: Tuesday, February 20, 2024, 2:00 PM - 3:00 PM
Where: G O Jones Building, Room 610, Mile End
Speaker: Bruno von Bruening
Information-Theory-based Electron Density Partitioning and Machine Learning for Force Field Development
High Accuracy Force Fields enable the modelling of molecular crystals, biomolecular systems, or host-guest complexes. In fact, diligent ab-initio parametrization allows to achieve CCSD(T) accuracy and DFT is out beaten - particularly in respect to intermolecular interaction.
Intermolecular interaction energies and in particular the anisotropic steric information arise to large extents from electrostatics. The electrostatic potential can be modelled by distributed multipoles derived from the quantum mechanical electron density. Among different postulated density partitioning schemes, the Iterative Stockholder Algorithm motivated by Information Theory is particularly powerful and will be presented in detail.
Next to the known physical forms, Machine Learning allows us to fit complicated quantum mechanical features of intermolecular potentials. In accordance with current trends to hybrid physics-machine learning schemes, we will rely on multipole-centered functional forms featuring parameters derived from the electronic density supported by Machine Learning terms. Among different strategies for learning the change of multipoles with conformation E(n)-Equivariant Graph Neural Networks will be highlighted.