Multi-task scattering-model classification and parameter regression of nanostructures from small-angle scattering data

Abstract

Machine learning (ML) can be employed at the data-analysis stage of small-angle scattering (SAS) experiments. This could assist in the characterization of nanomaterials and biological samples by providing accurate data-driven predictions of their structural parameters (e.g. particle shape and size) directly from their SAS profiles. However, the unique nature of SAS data presents several challenges to such a goal. For instance, one would need to develop a means of specifying an input representation and ML model that are suitable for processing SAS data. Furthermore, the lack of large open datasets for training such models is a significant barrier. We demonstrate an end-to-end multi-task system for jointly classifying SAS data into scattering-model classes and predicting their parameters. We suggest a scale-invariant representation for SAS intensities that makes the system robust to the units of the input and arbitrary unknown scaling factors, and compare this empirically to two other input representations. To address the lack of available experimental datasets, we create and train our proposed model on 1.1 million theoretical SAS intensities which we make publicly available. These span 55 scattering-model classes with a total of 219 structural parameters. Finally, we discuss applications, limitations and the potential for such a model to be integrated into SAS-data-analysis software.