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 64: Topical Session: Data Driven Materials Science - Materials Data Management (joint session MM/CPP)

CPP 64.3: Vortrag

Mittwoch, 18. März 2020, 11:00–11:15, BAR 205

Benchmarking neural networks on sequence-determined polymer transport through lipid membranes — •Marco Werner¹, Yachong Guo², and Vladimir Baulin³ — ¹Institut Theorie der Polymere, Leibniz-Institut für Polymerforschung Dresden, Germany — ²National Laboratory of Solid State Microstructure, Department of Physics, Nanjing University, China — ³Departament d’Enginyeria Química, Universitat Rovira i Virgili, Tarragona, Spain

We consider the transport of amphiphilic polymers through lipid membranes by passive diffusion as a function of the sequence of hydrophilic and hydrophobic building blocks. Massively parallel Rosenbluth sampling of polymer conformations is performed to estimate polymer translocation times through a membrane for all 2^N sequences and chain lengths N≤16. Our results confirm that smallest translocation times are found for polymers with balanced fraction of hydrophilic and hydrophobic units, and containing short blocks. Sequence-complete databases deliver an important ground truth for benchmarking machine-learning models against training data restrictions and biases. We demonstrate that multi-layer artificial neural networks show remarkable generalization performance when restricting the training data to relatively narrow windows of translocation times. The results indicate that relevant sequence patterns and their physical effect are approximated based on the restricted training set, however, accuracy drops towards unexplored corners in sequence space.