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 32.10: Vortrag
Montag, 16. März 2020, 18:00–18:15, HÜL 186
Adversarial Reverse Mapping of Equilibrated Condensed-Phase Molecular Structures — •Marc Stieffenhofer1, Michael Wand2, and Tristan Bereau1 — 1Max Planck Institute for Polymer Research, Ackermannweg 10, 55128 Mainz — 2JGU Mainz, Institute of Computer Science, Staudingerweg 9, 55128 Mainz
Coarse-grained molecular dynamics simulations circumvent prohibitively long equilibration times by averaging out degrees of freedom. However to consistently link the scales a reverse mapping scheme is needed to reintroduce higher resolution details. Traditional schemes propose a rough coarse-to-fine mapping and rely on further energy minimization and molecular dynamics simulations to re-equilibrate the system. In this study we introduce DeepBackmap: A deep neural network based approach to predict equilibrated molecular structures without running molecular dynamics simulations. We use generative adversarial networks to learn the Boltzmann distribution from training data and realize reverse mapping by using the coarse-grained structure as a conditional input. Our method can be used for condensed-phase systems of arbitrary sizes. The model uses only local information and reconstructs the atomistic structure autoregressively. We test our method on syndiotactic polystyrene molecules and show that the model trained in a melt shows remarkable transferability to the crystalline phase. The learned local correlations appear to be temperature independent indicating a separation of the scales.