Coherent Surface Scattering Imaging reconstruction with Pytorch autograd and multislice forward model

Coherent Surface Scattering Imaging (CSSI) brings together conventional imaging techniques, such as lensless X-ray Coherent Diffraction Imaging (CDI), transmission-geometry ptychography and laminography, and a surface sensitive technique of Grazing Incidence Small Angle X-ray Scattering (GISAXS). CSSI will be the feature technique of the new 9-ID beamline of the Advanced Photon Source Upgrade. For image reconstruction purposes, we need a physical model that can reproduce complex GISAXS scattering patterns, such as dynamical scattering fringes that are observed near sample horizon. Distorted Wave Born Approximation (DWBA) cannot reproduce such dynamical fringes near the critical angle of total external reflection. On the other hand, multislice simulations compute wave propagations through objects one 2D slice at a time along the beam direction and are already widely used in transmission geometry electron microscopy and X-ray transmission experiments. The multislice formalism could also be applied to CSSI reflection-geometry setups and can successfully reproduce dynamical scattering phenomena near critical angles. This multilice formalism has been implemented in fast performing GPU codes which can do forward calculations of tens of micron-scale surface patterns in a few seconds. Here, it is discussed how backward propagation of the multislice model and Pytorch auto-differentiation tool enable us to do 3D image reconstructions in the form of CSSI-CDI using a single scattering image.