Regensburg 2025 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
CPP: Fachverband Chemische Physik und Polymerphysik
CPP 15: Poster Session I
CPP 15.7: Poster
Monday, March 17, 2025, 19:00–21:00, P4
Simulation-Based Neural Network with Embedded Prior Knowledge for Predicting Morphological Parameters in GISAXS — •Shachar Dan1, Eldar Almamedov2, Matthias Schwartzkopf1, Sven-Jannik Wöhnert1, Andre Rothkirch1, Yufeng Zhai1, Jose I. Robledo4, Volker Skwarek2, and Stephan V. Roth3 — 1DESY, Notkestraße 85, D-22607 Hamburg — 2HAW, Berliner Tor 5, D-20099 Hamburg — 3KTH, Teknikringen 56 SE-10044 Stockholm — 4FZ-Jülich, Wilhelm-Johnen- Straße D-52428 Jülich
In-situ grazing-incidence small-angle X-ray scattering (GISAXS) is a powerful technique for analyzing nanoscale structures, yet its interpretation is challenging due to the inverse problem caused by phase information loss. Advances in simulation software and deep learning techniques have opened the door to the idea of using simulations, which can now be generated more efficiently in diverse configurations, to train neural networks (NNs). However, simulations often fail to fully represent experimental data, creating a significant sim-to-real gap. In our work, we tackle this challenge by embedding prior knowledge about the system into the NN training process. By incorporating constraints based on this knowledge, we train models on simulations and apply them to experimental data, enabling reasonable predictions of morphological parameters such as cluster radii, inter-cluster distances, and grain size distributions. This approach aims to accelerate material characterization at the nanoscale and provide a portable and efficient counterpart to traditional methods.
Keywords: GISAXS; Deep Learning; Neural Networks; Simulations