Where to next? Navigating neutron diffraction experiments with machine learning
Neutron diffraction experiments directly measure complex magnetic ordering in a material, which can be important for applications in magnetocaloric refrigeration and topological materials.
However, because there are currently only 11 neutron user facilities globally, beam time is limited and highly competitive. Traditionally, neutron diffraction experiments are scheduled ad hoc, probing material behavior in incremental steps, which can waste valuable beam time.
McDannald et al. developed ANDiE, the Autonomous Neutron Diffraction Explorer, which improved measurement efficiency by a factor of five. ANDiE reproduced the results for a well-studied material and discovered a previously unknown sharp transition behavior in another material.
The user inputs prior estimates about the neutron diffraction pattern and transition parameters, then ANDiE guides the experiment until it outputs the most likely type of magnetic transition and transition temperature.
“One major constraint on the temperature of the experiment is that data is only collected as the sample is warmed,” said author Austin McDannald. “This is because the measurement of the transition temperature depends on whether the sample is warming or cooling.”
To account for this limitation, ANDiE considers the model uncertainty as the prediction extrapolates to higher temperatures. The next measurement is taken when the model uncertainty is greater than some factor multiple of the estimated measurement uncertainty.
“We call this the ‘Bravery factor’, as it controls how uncertain the user is willing to let the extrapolation become before a measurement is taken, and therefore controls how ambitiously ANDiE explores the space,” said McDannald.
ANDiE has already been implemented at neutron diffraction facilities, and the researchers plan to adapt the technique for similar experiments.
Source: “On-the-fly autonomous control of neutron diffraction via physics-informed bayesian active learning,” by Austin McDannald, Matthias Frontzek, Andrei Savici, Mathieu Doucet, Efrain Rodriguez, Kate Meuse, Jessica Opsahl-Ong, Daniel Samarov, Ichiro Takeuchi, William D Ratcliff, and A. Gilad Kusne, Applied Physics Reviews (2022). The article can be accessed at https://doi.org/10.1063/5.0082956 .
This paper is part of the Autonomous (AI-driven) Materials Science Collection, learn more here .