Machine learning assisted masking of parasitic signals in Bragg coherent diffraction imaging

Bragg coherent diffraction imaging (BCDI) is a lens-less technique capable of imaging the strain in a particle in the size range from 20 nm up to several micrometres. This indirect measurement technique, used in X-ray synchrotrons or free-electron lasers all over the world, requires an inversion step using iterative algorithms in order to recover the real-space complex object encoding the particle shape and deformation field. However, artefacts such as scattering peaks called `aliens' from nearby particles can affect the accuracy of the final reconstruction and require meticulous and time-consuming manual masking of the raw data. This becomes problematic for BCDI reconstructions during an experiment and/or for large volumes of data. Here, we explore the potential of machine learning, and specifically clustering techniques, to speed up this procedure while keeping the maximum spatial resolution of the object reconstruction. We also provide a user-friendly Python Jupyter notebook program available on Github.