Dresden 2020 – wissenschaftliches Programm

Die DPG-Frühjahrstagung in Dresden musste abgesagt werden! Lesen Sie mehr ...

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

CPP: Fachverband Chemische Physik und Polymerphysik

CPP 32: Condensed-matter simulations augmented by advanced statistical methodologies (joint session DY/CPP)

CPP 32.8: Vortrag

Montag, 16. März 2020, 17:30–17:45, HÜL 186

Learning effective collective variables for biasing via t-distributed stochastic neighbor embedding — •Omar Valsson¹ and Jakub Rydzewski² — ¹Max Planck Institute for Polymer Research, Mainz, Germany — ²Institute of Physics, Faculty of Physics, Astronomy and Informatics, Nicolaus Copernicus University, Torun, Poland

A common strategy to overcome the time scale problem of molecular dynamics (MD) simulations is to employ collective variable (CV) based enhanced sampling methods [1]. However, the efficiency of such approaches depends critically on the quality of the chosen CVs that must describe all the slow degrees of freedom. While physical and chemical intuition has proven generally successful in achieving this, there is a growing need for methods that can automatically find good CVs. An appealing option to accomplish this is to use ideas from the field of machine learning.

Here we show how dimensionality reduction via t-distributed stochastic neighbor embedding (t-SNE) [2] can be employed to learn effective CVs for biasing. We have implemented t-SNE directly in the PLUMED plug-in, which allows us to use it directly in numerous MD codes.

[1] O. Valsson, P. Tiwary, and M. Parrinello, Ann. Rev. Phys. Chem. 67, 159 (2016)

[2] L. van der Maaten and G. Hinton, J. Mach. Learn. Res. 9, 2579 (2008)