Using Machine Learning methods to obtain the properties of water

DIPC Seminars

Alexandre Reily Rocha, Instituto de Física Teórica (IFT), Universidade Estadual Paulista (UNESP), São Paulo, Brazil
Hybrid Seminar: Donostia International Physics Center
Add to calendar
Subscribe to Newsletter
Using Machine Learning methods to obtain the properties of water Water is one of the most fundamental substances know to us, however some properties are not yet understood. This comes from the interplay between long and short range interactions, which give rise to a plethora of phenomena in water. First principles calculations are a powefull tool to describe different effects in water. Nevertheless, the high computational cost is a problem for systems with a large number of molecules. An alternative is to use neural networks trained with DFT (density functional theory) calculations. With this metodology we can describe the molecular interactions of small systems using DFT and apply them to much larger systems, and for longer timescales. Thus, using molecular dynamics (MD) with a neural-network trained force field we describe a number of properties of water - both dynamical and static - comparing different exchange and correlation potentials. We show that even simple PBE-type potentials with van der Waals corrections can yield very good results. At the end of my talk I will also discuss methods to obtain the smallest possible data set used for training. Host: Thomas Frederiksen Zoom: