view article
Figure 2
A schematic diagram of the network architecture (see Section 2 of the supporting information for more details). The discriminators were fed GT and generator output samples. They were trained to label the former as real and the latter as fake. Both inputs to the discriminators were paired with their respective clean scattering data (i.e. without the artefacts). The generators were trained adversarially, receiving the scattering data with artefacts as input into both IFF(Q) and ISRO(Q) U-Net generators, which encoded and decoded this into our desired output. After one full iteration of training for the discriminators and generators, an additional training stage occurred where the IFF(Q) and ISRO(Q) generated outputs were multiplied, and the error between this and the clean scattering data was used to train the generators.

IUCrJ
Volume 11| Part 1| January 2024| Pages 34-44
ISSN: 2052-2525
https://doi.org/10.1107/S2052252523009521