Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
CPP: Fachverband Chemische Physik und Polymerphysik
CPP 46: Poster II
CPP 46.24: Poster
Donnerstag, 12. März 2026, 09:30–11:30, P5
Physics-Informed Neural Inference with Multilayer DWBA Modeling for Quantitative GISAXS — •Özüm Emre Asirim1,2, Yufeng Zhai1, Marina Tropmann-Frick2, and Stephan V. Roth1,3 — 1Deutsches Elektronen-Synchrotron DESY FS-SMA Notkestr. 85 22607 Hamburg Germany — 2HAW Hamburg Berliner Tor 7 20099 Hamburg Germany — 3KTH Royal Institute of Technology, SE-100 44 Stockholm, Sweden
Quantitative GISAXS analysis is often constrained by the complexity of multilayer scattering, film-nanoparticle coupling, and the limited availability of physically interpretable synthetic datasets for training modern inference models. We address these challenges with a physics-informed neural framework that integrates a compact multilayer DWBA formulation with a hybrid CNN-Transformer architecture. The forward model captures essential film and nanoparticle scattering through per-layer incoherent DWBA terms, normalized form factors, structure factors, and DWBA-based field enhancement, yielding a reduced and physically coherent parameter space suitable for inversion. Measurement variability is modeled through noise, detector effects, and the ability to generate time-series data for evolving structures. Leveraging this parameterization, the CNN-Transformer network jointly extracts local scattering motifs and long-range reciprocal-space correlations, enabling accurate recovery of film- and particle-level parameters. Validated through forward reconstruction using the same physics model, the workflow provides a fast, interpretable, and physically grounded solution for high-throughput GISAXS analysis.
Keywords: GISAXS; DWBA; Beamline; Machine Learning; Nanostructure