Deep generative priors for ptychography

Self-supervised neural priors make electron ptychography faster, sharper, and more robust to noise.

projects

Papers: Deep generative priors for robust and efficient electron ptychography, arXiv (2025) · Accelerating iterative ptychography with an integrated neural network, Journal of Microscopy (2025)

Electron ptychography recovers the full complex transmission function of a sample, potentially at deep sub-ångström resolution, from a 4D-STEM dataset. In practice, conventional pixel-based reconstructions are sensitive to noise and need careful hand-tuned regularization, especially for thick samples reconstructed in 3D with multislice methods.

Rather than directly learning the pixelated object and probe, we use deep generative priors: neural networks that parameterize the object and probe inside the same automatic-differentiation multislice forward model. The networks are trained per-dataset and self-supervised, so no training data is required, and their inductive biases regularize the reconstruction. Compared with state-of-the-art iterative and ML baselines, this gives a >10× speedup and 40% better resolution, with improved robustness at low dose and better depth regularization.

Deep generative priors (DGPs) for the probe and object feed a differentiable multislice model, alongside auxiliary parameters such as scan positions, descan, and beam tilt. Everything is optimized together against the measured data.

The benefit is clearest on real, low-dose data. On a public dataset of the MOSS-6 metal-organic framework, recorded at 100 e⁻/Ų, using DGPs for both the object and the probe reduces noise across a wide range of spatial frequencies without attenuating the signal. The information limit improves from 1.98 Å to 1.57 Å.

Multiple reconstructions of the MOSS-6 metal-organic framework from the same experimental dataset. (a) Object phase with DGPs generating both the object and the probe, and (b) its FFT; the dashed circle marks 1.7 Å. (c, d) Pixelated object and probe. (e, f) DGP object with a pixelated probe. (g, h) Pixelated object with a DGP probe. Insets magnify the red boxes.

The same framework now underpins projects in tomography and crystallography, and the implementation lives in quantEM. You can run the code yourself in the quantEM tutorial notebook.

Interested in how multislice ptychography works? Try the interactive ptychography demo.