Regensburg 2022 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

CPP: Fachverband Chemische Physik und Polymerphysik

CPP 12: Poster 1

CPP 12.22: Poster

Monday, September 5, 2022, 18:00–20:00, P1

Propagation of learned sequence patterns to larger chain length using TransEncoder neural networks — •Huzaifa Shabbir, Marco Werner, and Jens Uwe Sommer — Leibniz Institute for Polymer Research Dresden

In this work, we investigate artificial neural networks that are capable of learning and transferring hidden variables in chemical sequences from a small sequence length to a larger one. Patterns in the relation between the hydrophilic/hydrophobic sequence of a copolymer and its free energy of interaction with a lipid membrane have been learned with the aid of AutoEncoder neural networks, which were employed to translate between these two properties (TransEncoder)[1]. We demonstrate that the latent space parameters learned by the TransEncoder allow for a physical interpretation of the contributions to the free energy. Furthermore, the learned patterns for a smaller sequence length can be transferred towards a higher sequence length of our interest, which not only significantly reduces the number of training examples required but also increases the accuracy in comparison to the training for individual polymer sequence length. We investigate the computational efficiency and the convergence of learned patterns when multiple chain lengths are addressed at once.

[1] M. Werner, ACS Macro Letters 10, 1333 (2021).