Physics-Aligned Self-Supervised Learning for Scientific Imaging — Bashir Kazimi, FZJ | HAICON26

Описание: Speaker: B. Kazimi
Affiliation: Forschungszentrum Jülich (IAS-9) / RWTH Aachen, Germany
Date: June 10, 2026 Time: 9:28–9:42am Session: Session 2b – Domain-informed Methods

Recorded at the Helmholtz AI Conference (HAICON26), June 9–11, 2026. More info: https://haicon.cc/conference-program/

About this talk: Physics-Aligned Self-Supervised Learning for Scientific Imaging B. Kazimi 1, S. Sandfeld 1 1 Forschungszentrum Jülich (IAS-9) / RWTH Aachen, Germany

This work studies physics-aligned augmentations for self-supervised learning in electron microscopy and 4D-STEM. DINOv2 with physics-aligned augmentations improves classification accuracy from 66.70% to 76.67% on NFFA, and reduces orientation error from 9.85° to 5.60° on 4D-STEM, with improved spectral utilization and robustness.