Physics-Informed Multimodal Learning for Phase Retrieval from Synchrotron Data

Synchrotrons have played an instrumental role in unravelling atomistic to mesoscale details for a variety of material systems. It is through the culmination of several scattering techniques that equipped us with a lens to investigate these surface, bulk and atomistic features of dynamical systems. However, the advancements are limited by two key major aspects, one is the curation of high volumes of data for further analysis and the other is online computation necessary to construct the real space from reciprocal space data. The latter has been addressed by utilization of several phase learning involving both conventional optimization and deep learning (DL) methods. Out of them tools based on DL have shown that they are scalable on the modern edge computing devices. However, they require multiple training cycles and are often not complemented with relevant physics. This makes it difficult to validate the complex-uninterpretable neural network models. We circumvent this by utilizing a multimodal dataset from Small Angle X-ray Scattering and X-ray Photoelectron Spectroscopy. We use PMMA/PA as our representative system for which we first test our workflow on a synthetic dataset and implement the same for a real-time data stream. We further anticipate that this workflow could be utilized to simulate much more physical datasets via noise-learning from real-time data.