Dresden 2026 – scientific programme

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

DS: Fachverband Dünne Schichten

DS 20: Poster

DS 20.34: Poster

Thursday, March 12, 2026, 18:30–20:30, P2

Domain Transfer from Simulation to Experimental Neutron and X-ray Reflectivity Data Using Probabilistic Generative Models — •Jeyhun Rustamov¹, Ritz Aguilar¹, Vedhas Pandit¹, Nico Hoffmann², and Jeffrey Kelling¹ — ¹Helmholtz-Zentrum Dresden-Rossendorf, Dresden, Germany — ²Helm & Walter IT-Solutions, Dresden, Germany

Machine learning (ML) models trained on simulations often fail to generalize to experimental data in neutron and X-ray reflectivity analysis. Furthermore, determining thin film parameters from reflectivity curves is an inherently ambiguous inverse problem. To address this, we employ conditional normalizing flows (cNFs) with a β-Variational Autoencoder (β-VAE) embedding network to learn the full distribution of physical parameters instead of single estimates.

To further improve performance on experimental data, we systematically explore three strategies for bridging the simulation-experiment domain gap: fine-tuning with labeled experimental data, utilizing generative models to create realistic synthetic data, and a novel physics-informed method. The proposed method incorporates a differentiable forward function based on the kinematic approximation to guide the generation of sample parameters via physics-informed loss during bidirectional training of cNFs.

Our approach distinctly leverages unlabeled experimental reflection data to address cNF performance challenges on real-world reflectivity measurements. Additionally, this methodology can be broadly applied to other inverse problems with similar domain-transfer challenges.