Regensburg 2019 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
CPP: Fachverband Chemische Physik und Polymerphysik
CPP 44: Condensed-matter simulations augmented by advanced statistical methodologies (joint session DY/CPP)
CPP 44.6: Vortrag
Mittwoch, 3. April 2019, 16:30–16:45, H20
Deep Learning for Multiscale Simulations of Soft Matter Systems — •Marc Stieffenhofer1, Tristan Bereau1, and Michael Wand2 — 1Max Planck Institute for Polymer Research, Ackermannweg 10, 55128 Mainz — 2Institute of Computer Science, Johannes Gutenberg University Mainz
The great success of Deep Neural Networks (DNNs) is based on their ability to learn and extract descriptive features directly from training data and to build a hierarchical, abstract representation of the input that takes multiple length scales into account.
Such multiscale representations can also be found in soft matter systems where many physical phenomena and properties are governed by a large range of different length- and timescales.
In this work, we explore links between multiscale representations of DNNs and multiscale simulations of soft matter systems. The main focus is to investigate if DNNs can be used to link distribution functions generated at different resolutions. We have applied DNNs to the backmapping of coarse-grained molecular configurations to higher-resolution representations. This requires to reproduce the fine-grained statistics that match the coarse-grained representation.
Our model relies on 3D convolutional neural networks that are trained to generate molecular equilibrium structures. The training of the model is based on the generative adversarial approach and results are discussed for a system of octane molecules.