Feature Engineering for Small-Angle Scattering Model Selection

While many soft-matter characterization techniques can be unambiguously interpreted to yield information about chemical and material structure, small-angle neutron and X-ray scattering (SANS and SAXS) data must be interpreted via a library of physical and phenomenological models. Complicating this task is the unavoidable phase problem, which causes scattering patterns to be non-unique and makes model selection a non-trivial task. Here we present our efforts in developing shallow- and deep-classifiers which, given a scattering dataset, suggest applicable models to the user. In particular, we will focus on our efforts in feature vector engineering i.e., the optimization of input parameters in order to maximize classification efficiency. We show that simple data transformations greatly increase our classification efficiency over a naïve model, allowing us to achieve greater than 99 % top-3 accuracy in the model-selection task. More broadly, these optimized feature vectors will enhance machine learning models for tasks other than model selection (e.g., error detection, automated experimentation, on-the-fly analysis). The ultimate goal of this project is to formalize feature vector design for small-angle scattering, thereby enabling the creation of bespoke machine-learning models.