The NIST Autonomous Formulation Laboratory: Virtual Scattering Instruments for Active Learning Agent Development

Machine-learning (ML) is revolutionizing the way we do material science with ML models that are discovering new polymer chemistries, designing syntheses, guiding measurements, and analyzing data, all with unprecedented accuracy and speed. Despite these advances, many challenges remain in creating the next generation of polymer ML models. A key limitation for polymer science focused models is the lack of high-quality, well-tagged measurement data in sufficient amounts for supervised training. The generation of measurement data for soft materials is hampered by limited polymer availability, tedious sample preparation, slow instruments, and involved data processing. While synthetic data can help alleviate this data scarcity, it is challenging to produce data that correctly accounts for instrument artifacts and resolution effects. Beyond the instrumental effects, mimicking the behavior of real soft-material samples requires accounting for impurities, processing effects, and other non-idealities that go beyond simple analytical models. In this talk we will describe our attempts to overcome these challenges with the development of a virtual small-angle scattering (SAS) instrument. This goal of our framework is to produce high-quality SAS data as a function of input parameters (e.g., sample composition, processing parameters, instrument configuration) that can be used for model training or ML agent development. We will specifically highlight the design of the virtual instrument and how it can be used to tune active learning agents. Finally, we will demonstrate how these tuned agents are performant in guiding real small-angle scattering experiments.