Regensburg 2022 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
CPP: Fachverband Chemische Physik und Polymerphysik
CPP 14: Emerging Topics in Chemical and Polymer Physics, New Instruments and Methods
CPP 14.3: Vortrag
Dienstag, 6. September 2022, 10:00–10:15, H39
Transferable hidden variables in sequence space learned by transencoder neural networks — •Marco Werner — Institut Theorie der Polymere, Leibniz-Institut für Polymerforschung Dresden, Germany
The relation between chemical sequences and the properties of polymers is investigated using artificial neural networks with a bottleneck layer of neurons. By training such AutoEncoder networks to translate between sequence and property (TransEncoder1), one may identify variables that control the physical relationship behind. Here, networks were trained to predict the effective free energy landscape of a copolymer interacting with a lipid membrane depending on its sequence of hydrophilic and hydrophobic monomers. TransEncoders that were split into separate encoder-decoder channels have learned to decompose the free energy into independent components that were physically meaningful. For instance, they reflect theoretical concepts such as solutions of the Edwards equation. Sequence-complete data sets for training were obtained via Rosenbluth sampling of single chains in a given density field. It is demonstrated that once the sequence patterns were learned based on the large data set for chain length N=14, a small number of ∼20 examples was sufficient to transfer-learn the prediction to a more detailed simulation model with explicit lipids and solvent (accuracy 0.5kBT). The results open a perspective to physics-informed inverse searches, for instance, for copolymer sequences leading to the smallest translocation time through a membrane. [1] M. Werner, ACS Macro Lett. 10, 1333 (2021).