We propose a framework for utilizing machine learning-based models to derive information from scattering data that human users cannot currently determine quickly without the help of computationally expensive data analysis. We train our models to not only identify morphology, but also inform the quantitative value of structural parameters without requiring expensive optimization procedures. Leveraging physical knowledge from the known forward models, we develop novel algorithms that are designed to accurately extrapolate to different structural parameter ranges, unlike traditional machine learning models which are designed to be predictive over the same distribution as the training data. We discuss how to curate a simulated dataset for training our machine learning models to enable high quality predictions on experimental data. The machine learning software we develop can be integrated with existing scattering analysis software such as SasView. Our machine learning framework can help enable real-time data analysis for experiments, increase the efficiency of expert and facility time, and democratize scattering technology.

Machine learning-enabled inverse analysis of scattering data

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.