Jeffrey Kelling

Jeffrey Kelling

Group leader

Xray Diffraction

We are currently developing a one-step reconstruction method for inverting SAXS data during experiments at synchrotrons like DESY or our HIBEF beamline at EuropeanXFEL. Due to the loss of the phase information, there is often more than one solution for each diffraction pattern. The conditional Invertible Network, a generative model, is able to reconstruct all possible solutions in a matter of milliseconds. The network is trained on perturbed simulated data only with a data-driven negative log-likelihood loss and is able to reconstruct most experimental data.

The library nfPhasing expands the data loss by a physics-driven loss that integrates knowledge about the underlying scattering processes.

The general idea is that a conditional Normalising Flow is learning a mapping from experimentally acquired diffraction pattern(s) I and some prior distribution to the predictive posterior distribution of electron densities u. The neural network is trained by a data-driven objective as well as a Physics-based loss. The former allows for very fast inference (i.e. reconstruction) on data similiar to our training data while the latter enables reconstruction of objects that are out-of-distribution to the training data.

nfPhasing is covering the following modalities:

  • 1d/2d Small-angle X-ray scattering (at Grazing Incidence)
  • 2d Ptychography
  • 2d Holography / Phase Contrast Imaging

Team

  • Ritz-Ann Aguilar
  • Jeyhun Rustamov
  • Kristin Tippey

Publications

  • Achilles, S., Ehrig, S., Hoffmann, N., Kahnt, M., Becher, J., Fam, Y., Sheppard, T., Brückner, D., Schropp, A., Schroer, C. (2022). GPU-accelerated coupled ptychographic tomography. Proc. SPIE 12242, Developments in X-Ray Tomography XIV, 122420N. paper

  • Zhdanov, M., Randolph, L., Kluge, T., Nakatsutsumi, M., Gutt, C., Ganeva, M., Hoffmann, N. (2022). Amortized Bayesian Inference of GISAXS Data with Normalizing Flows. Machine Learning and the Physical Sciences workshop @ NeurIPS 2022. paper

  • Liu, Y., Shao, Z., Teng, Y., Hoffmann N. (2021). NAM: Normalization-based Attention Module. ImageNet PPF @ NeurIPS 2021. paper

  • Liu, Y., Shao, Z., Hoffmann, N. (2021). Global Attention Mechanism: Retain Information to Enhance Channel-Spatial Interactions. arXiv:2112.05561v1. paper