Brookhaven AIMS Series

Data reconstruction is the key step to interpreting the measurement signal into structural information in synchrotron X-ray experiments. It is always challenging due to the complexity of the measurement modalities and data conditions. I will present my developments in using deep neural networks for data reconstruction in synchrotron X-ray measurements. These works are model-based learning, in which no training data and training process are required. My first development case is a 2D image reconstruction algorithm GANrec (Generative Adversarial Network reconstruction algorithm). It was applied to X-ray tomographic reconstruction and holographic phase retrieval. It showed improved reconstruction quality and accuracy for extreme data conditions. The other case is a 1D signal reconstruction process with 1D CNN (Convolutional Neural Network) inverse solver. It was tested for coded-aperture data reconstruction. The results showed less noise and reduced error compared with the classical methods.