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

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

MM: Fachverband Metall- und Materialphysik

MM 17: Computational Materials Modelling - Methods II

MM 17.2: Vortrag

Montag, 16. März 2020, 17:15–17:30, IFW D

New basis functions and optimization algorithms for Variationally Enhanced Sampling — •Benjamin Pampel and Omar Valsson — Max Planck Institute for Polymer Research, Mainz, Germany

Variationally Enhanced Sampling is an advanced sampling method for molecular dynamics simulations based on a variational principle. A bias potential is construced by minimizing a convex functional, to obtain thermodynamic and kinetic information of rare event systems.

So far mostly orthogonal polynomials have been used as basis functions for the expansion of the bias potential. While there have been proposals for alternatives, this was for specific problems and not with a general evaluation of the performance of the method in mind. Therefore, the most efficient choice of basis functions is an open question.

Another important ingredient of the VES method is the employed optimization algorithm. While the Bach’s averaged stochastic gradient decent used so far has generally performed fine, more recent algorithms from the field of machine learning might perform better.

To answer these two open questions, we first implemented and evaluated the performance of new sets of basis functions, including Gaussians and cubic splines, by testing them on various systems. We find the usage of Daubechies Wavelets favorable for many applications.

For the optimization, besides implementing some of the popular algorithms like Adagrad or Adam, we also propose our own modifications. Here the conclusions are less clear, as only in few cases the Adam algorithm is able to outperform Bach’s algorithm.