Dresden 2026 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
DS: Fachverband Dünne Schichten
DS 21: Optical Analysis of Thin Films
DS 21.4: Vortrag
Freitag, 13. März 2026, 10:30–10:45, REC/C213
VIPR: A Modular Machine Learning Framework for Inverse Problems with Application to Reflectometry — •Sascha Creutzburg1,6, Jeyhun Rustamov2, Alexandros Koutsioumpas4, Jens Bornschein1,6, Marina Ganeva4, Stefan Häusler4, Bernd Helm1,6, Alexander Hinderhofer5, Mykhailo Levytskyi4, Valentin Munteanu5, Robert Juzak1,6, Vedhas Pandit2, Frank Schreiber5, Jeffrey Kelling2,3, and Nico Mothes1,6 — 1Helm & Walter IT-Solutions GmbH, Dresden — 2Helmholtz-Zentrum Dresden-Rossendorf — 3Chemnitz University of Technology — 4Jülich Centre for Neutron Science, Heinz Maier-Leibnitz Zentrum — 5University of Tübingen — 6Saxony.AI
Ambiguous inverse problems are ubiquitous in experimental physics, where multiple parameter configurations can reproduce the same measurement and classical iterative approaches are often too slow for online analysis. We present the VIPR framework, a modular machine-learning approach for such problems in thin-film reflectometry. Its reflectometry plugin for X-ray and neutron reflectometry combines three approaches: (i) Reflectorch for fast parameter estimation, (ii) prior-amortized neural posterior estimation, and (iii) neural spline flows to capture multimodal posteriors and reveal alternative plausible structures. VIPR features streaming mode for near-real-time beamline feedback and is deployed at Heinz Maier-Leibnitz Zentrum for routine analysis. The framework integrates seamlessly into existing workflows and can be extended to other inverse problems via its plugin architecture. Code: https://codebase.helmholtz.cloud/vipr/vipr-framework
Keywords: Machine Learning; Reflectometry; Bayesian Inference; Normalizing Flows; Inverse Problems