Application of generative physics-informed neural networks for phase retrieval in the Fresnel regime

• Grant number: 26/13819-8
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: August 01, 2026
• End date: July 31, 2027
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computational Mathematics
• Principal Investigator: Daniel Roberto Cassar
• Grantee: Mariana Melo Pereira
• Host Institution: Centro Nacional de Pesquisa em Energia e Materiais (CNPEM). Campinas , SP, Brazil
• Associated research grant: 24/00989-7 - Research Center of Molecular Engineering for Advanced Materials (CEMol), AP.CEPID

Abstract

X-ray phase contrast in the Fresnel regime is a key technique for imaging weakly absorbing samples and is widely used at several synchrotron facilities around the world, including Sirius, the fourth-generation particle accelerator located at the Brazilian Center for Research in Energy and Materials (CNPEM) in Campinas, São Paulo. In this regime, scattering effects are significant, making the wave phase a fundamental component of the observed contrast. However, during experimental acquisition, phase information is not measured directly, giving rise to the phase retrieval problem, which is classified as an inverse and ill-posed problem in which this information must be inferred from the intensities recorded by the detector. Conventional algebraic and iterative methods currently employed by the Scientific Computing Group (GCC) at Sirius present limitations due to the presence of noise, reliance on approximations, and the assumption of ideal experimental conditions, which are usually not satisfied. As an alternative, this project proposes the use of deep learning techniques to reconstruct phase and absorbance from simulated data through generative adversarial networks (GANs) and physics-informed neural networks (PINNs). It is expected that this approach will yield superior performance compared to traditional methods. A unique feature of the project is the evaluation of different network architectures to optimize the reconstructions. The ultimate goal (likely beyond this undergraduate research project) is to apply the trained model to data measured at the beamlines of the Brazilian Synchrotron Light Laboratory (LNLS). (AU)