DL4EM/physics-aligned-ssl

Physics-Aligned Self-Supervised Learning for Scientific Imaging

Pretrained ViT-B encoders from the GCPR 2026 paper Physics-Aligned Self-Supervised Learning for Scientific Imaging.

Code: DL4EM/physics-aligned-ssl

Five SSL methods (DINOv2, I-JEPA, MAE, SimCLR, VICRegL) were each pretrained on two electron-microscopy modalities under two augmentation regimes:

• *_domain — physics-aligned augmentations (T_phys): measurement-consistent symmetries plus acquisition-driven perturbations (noise, intensity variation, reciprocal-space scaling, diffraction tilt, ...).
• *_original — standard natural-image augmentations (T_orig): random crop, horizontal flip, blur, photometric perturbations.

Pretraining data:

• cem500k/ — real-space cellular EM (CEM500K subset, 10k images).
• 4dstem/ — simulated LiNiO2 4D-STEM diffraction patterns (Scheunert et al. subset, 10k patterns).

All encoders are single-channel (grayscale) ViT-B backbones.

Available models

Path	Method	Pretraining data	Augmentations
cem500k/dinov2_domain	dinov2	cem500k	physics-aligned
cem500k/dinov2_original	dinov2	cem500k	natural-image
cem500k/ijepa_domain	ijepa	cem500k	physics-aligned
cem500k/ijepa_original	ijepa	cem500k	natural-image
cem500k/mae_domain	MAE	cem500k	physics-aligned
cem500k/mae_original	MAE	cem500k	natural-image
cem500k/simclr_domain	simclr	cem500k	physics-aligned
cem500k/simclr_original	simclr	cem500k	natural-image
cem500k/vicregl_domain	vicregl	cem500k	physics-aligned
cem500k/vicregl_original	vicregl	cem500k	natural-image
4dstem/dinov2_domain	dinov2	4dstem	physics-aligned
4dstem/dinov2_original	dinov2	4dstem	natural-image
4dstem/ijepa_domain	ijepa	4dstem	physics-aligned
4dstem/ijepa_original	ijepa	4dstem	natural-image
4dstem/mae_domain	MAE	4dstem	physics-aligned
4dstem/mae_original	MAE	4dstem	natural-image
4dstem/simclr_domain	simclr	4dstem	physics-aligned
4dstem/simclr_original	simclr	4dstem	natural-image
4dstem/vicregl_domain	vicregl	4dstem	physics-aligned
4dstem/vicregl_original	vicregl	4dstem	natural-image

Usage

With the accompanying code (https://github.com/DL4EM/physics-aligned-ssl):

from em_ssl.hub import load_encoder

encoder = load_encoder("cem500k/dinov2_domain", repo_id="DL4EM/physics-aligned-ssl")

import torch
images = torch.randn(4, 1, 128, 128)   # grayscale EM crops in [0, 1]
features = encoder(images)

Without the codebase, each encoder.pt is a plain PyTorch checkpoint:

import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download("DL4EM/physics-aligned-ssl", "cem500k/dinov2_domain/encoder.pt")
ckpt = torch.load(path, map_location="cpu", weights_only=True)
state_dict = ckpt["encoder_state_dict"]     # ViT-B weights
print(ckpt["backbone"], ckpt["backbone_kwargs"])

Inputs are single-channel images normalised to [0, 1] (percentile normalisation was used during pretraining). Real-space EM models were trained on 128x128 crops; 4D-STEM models on 224x224 crops.

Citation

@inproceedings{kazimi2026physicsaligned,
  title     = {Physics-Aligned Self-Supervised Learning for Scientific Imaging},
  author    = {Kazimi, Bashir and Sandfeld, Stefan},
  booktitle = {DAGM German Conference on Pattern Recognition (GCPR)},
  year      = {2026}
}

About

Developed by the Deep Learning for Electron Microscopy (DL4EM) group at the Institute for Materials Data Science and Informatics (IAS-9), Forschungszentrum Jülich. For questions, please open an issue on the GitHub repository.