Morphology reconstruction from experimental small-angle x-ray scattering patterns by physics-aware neural network

In this work, we developed a new methodology that can reconstruct the morphology from experimental small-angle x-ray scattering (SAXS) patterns directly without modeling by using a physics-aware neural network, SAXSNN. By incorporating the scattering physics of x rays into the network, SAXSNN could be trained to capture the complex mapping between the SAXS patterns in reciprocal space and the corresponding morphologies in real space in an unsupervised way. We demonstrated the performance of SAXSNN on the experimental SAXS patterns of semicrystalline and amorphous polymers, i.e., hard-elastic isotactic polypropylene (iPP) films and plasticized poly(vinyl butyral) (PVB). The morphologies reconstructed by SAXSNN are well consistent with our existing knowledge of the morphology of iPP films and PVB. The developed methodology here allows us to rapidly predict the morphologies for any given SAXS pattern without any in-prior phase information and, thus, provides an intuitive understanding of the microstructures of the measured samples. A real-time feedback of the morphologies of measured samples to SAXS beamline users at modern synchrotron radiation light sources will be feasible in the near future.