In any experiment, some of the data collected are more useful for answering the scientific question at hand than the rest. If the useful data could be identified in advance, a significant amount of experiment time could be saved; unfortunately, the knowledge required for this identification is not gained until after the experiment is run. Active learning techniques can resolve this contradiction. These techniques operate cyclically, processing and analyzing experimental data and using the analysis in real-time to identify the most valuable data points to be collected next.
Neutron scattering techniques are inherently slow and are thus good candidates for active learning-based autonomous data collection. In addition, many neutron scattering techniques are fundamentally model-based. Examples include diffraction techniques, which measure shaped peaks in reciprocal space; neutron spin echo, which measures a damped sinusoid; or neutron reflectometry, for which data analysis relies on models of thin film structures. In each case, autonomous data collection can identify in real-time the data that will constrain the model parameters of interest in the least amount of measurement time.
Using simulated experiments on the CANDOR reflectometer, we determined that 2x - 5x speed-ups are realistic. We successfully devised a framework for efficient reflectometry data collection (manuscript in preparation ) and are currently working on the software implementation. Since the framework will collect reflectometry data out-of-order compared to traditional trajectory-based data acquisition, fundamental changes to the instrument control software are required. Those changes are tightly coordinated with the non-equilibrium data acquisition and FAIR data efforts.