An efficient ptychography reconstruction strategy through fine-tuning of large pre-trained deep learning model
Xinyu Pan,
Shuo Wang,
Zhongzheng Zhou,
Liang Zhou,
Peng Liu,
Chun Li,
Wenhui Wang,
Chenglong Zhang,
Yuhui Dong,
Yi Zhang
Affiliations
Xinyu Pan
Beijing Synchrotron Radiation Facility, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China; University of Chinese Academy of Sciences, Beijing 100049, China
Shuo Wang
Beijing Synchrotron Radiation Facility, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China; University of Chinese Academy of Sciences, Beijing 100049, China
Zhongzheng Zhou
Beijing Synchrotron Radiation Facility, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China; University of Chinese Academy of Sciences, Beijing 100049, China
Liang Zhou
Beijing Synchrotron Radiation Facility, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China; University of Chinese Academy of Sciences, Beijing 100049, China
Peng Liu
Beijing Synchrotron Radiation Facility, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China; University of Chinese Academy of Sciences, Beijing 100049, China
Chun Li
Beijing Synchrotron Radiation Facility, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China; Spallation Neutron Source Science Center, Dongguan, Guangdong 523803, China
Wenhui Wang
Beijing Synchrotron Radiation Facility, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China; Spallation Neutron Source Science Center, Dongguan, Guangdong 523803, China
Chenglong Zhang
Beijing Synchrotron Radiation Facility, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China; University of Chinese Academy of Sciences, Beijing 100049, China; Corresponding author
Yuhui Dong
Beijing Synchrotron Radiation Facility, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China; University of Chinese Academy of Sciences, Beijing 100049, China; Corresponding author
Yi Zhang
Beijing Synchrotron Radiation Facility, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China; University of Chinese Academy of Sciences, Beijing 100049, China; Corresponding author
Summary: With pre-trained large models and their associated fine-tuning paradigms being constantly applied in deep learning, the performance of large models achieves a dramatic boost, mostly owing to the improvements on both data quantity and quality. Next-generation synchrotron light sources offer ultra-bright and highly coherent X-rays, which are becoming one of the largest data sources for scientific experiments. As one of the most data-intensive scanning-based imaging methodologies, ptychography produces an immense amount of data, making the adoption of large deep learning models possible. Here, we introduce and refine the architecture of a neural network model to improve the reconstruction performance, through fine-tuning large pre-trained model using a variety of datasets. The pre-trained model exhibits remarkable generalization capability, while the fine-tuning strategy enhances the reconstruction quality. We anticipate this work will contribute to the advancement of deep learning methods in ptychography, as well as in broader coherent diffraction imaging methodologies in future.
