Automated structural analysis of small angle scattering data from common nanoparticles via machine learning

Billions of dollars have been invested in recent years to build up national scattering facilities around the world with more advanced configurations and faster data collection for small angle scattering (SAS), a technique that enables in situ structural analysis of nanoparticles (NP) under stringent sample environments. However, the interpretation of experimental SAS data is typically a slow process that requires significant domain expertise, leading to high-throughput scattering facilities such as synchrotron scattering centers collecting large quantities of data that may potentially be left unanalyzed. Here, we present a fast and data-efficient machine learning (ML) framework for identifying basic NP morphologies (spherical, cylindrical and discoidal geometries) and their corresponding structural parameters. The trained models take as input scattering curves with minimal pre-processing, and are able to identify morphology and structural dimensions from experimental curves with comparable accuracy to human experts. Critically, design choices that facilitate the practical application of ML models in scattering facilities are discussed, including ease of training, extrapolability outside of the parameter range of training data, and verifiability of predictions. The enhanced data analysis efficiency enabled by applying ML models to real-time in situ analysis of SAS data has the potential to revolutionize the utilization of synchrotron and neutron scattering facilities for probing nanostructures.