Paper accepted and presented at GCPR 2026

Our paper Physics-Aligned Self-Supervised Learning for Scientific Imaging was accepted at the DAGM German Conference on Pattern Recognition (GCPR 2026), held jointly with the Vision, Modeling, and Visualization Conference (VMV) from September 22 to 25, 2026 at the University of Siegen, Germany. I presented the paper as a poster at the conference.

The paper shows that the data augmentations used in self-supervised learning should follow the physics of the measurement rather than be copied from natural-image pipelines. We give a reproducible procedure for choosing such physics-aligned augmentations and evaluate it on real-space electron microscopy and 4D-STEM diffraction data across five self-supervised methods (DINOv2, SimCLR, MAE, VICRegL, I-JEPA).

• Preprint on arXiv
• Code on GitHub
• Pretrained models on Hugging Face

I also gave a talk on this work at HAICon26 in June.