23 May 2023
In neutron spectroscopy, neutrons are directed onto material samples and then collected by detectors and analysed. Some of the neutrons are scattered by the atoms of the sample, while others pass through the sample without interacting. Only the scattered neutrons contain information, such as how much energy the neutrons have absorbed or lost through the scattering process. The other neutrons generate so-called “noise”. In order to minimise the time spent measuring noise signals, as a first step measurements are carried out using a coarse grid in which the measurement points are evenly distributed. The algorithm then starts on its main task of using this initial data to identify areas in which further measurements would be beneficial. With each additional measurement point, the algorithm continues to supplement its own database and then autonomously decides on the next measurement location. On account of this feedback loop, this approach can also be described as 'active learning'.
Neutrons provide unique insights into the structure and dynamics of matter. This can only be achieved at large-scale research facilities, either at research reactors or specialised particle accelerators. For many years now, the number of available facilities in Europe has not been sufficient to meet the needs of research. Measurement time with neutrons is therefore both scarce and highly valuable. Methods to increase efficiency can help to make better use of available resources and reduce existing gaps.
The researchers were able to demonstrate the advantages of their approach in a real neutron experiment, on previously measured data sets, as well as numerous synthetic data sets. They were able to show that the available measurement time could therefore be used more efficiently compared to certain methods previously employed. At the core of the active learning algorithm is a mathematical distribution, the Gaussian curve. It describes a statistical distribution of data in the form of a bell curve. The researchers use this to target areas with informative signals. Other mathematical “tricks” allow areas with both strong and weak signals to be identified and differentiated from noise signals.
The short video compares two measurement approaches: on the left, the distribution of the measurement points solely following a schematic grid approach, and on the right, the active learning measurement method with hardly any measurement points lying in noise regions. (Copyright: Forschungszentrum Jülich)
Teixeira Parente, M., Brandl, G., Franz, C. et al.
Active learning-assisted neutron spectroscopy with log-Gaussian processes.
Nat Commun 14, 2246 (2023). https://doi.org/10.1038/s41467-023-37418-8
