Active learning-assisted neutron spectroscopy with log-Gaussian processes

TL;DR

This paper introduces an autonomous active learning method using log-Gaussian processes to optimize neutron spectroscopy experiments, significantly improving efficiency and reducing measurement time in TAS experiments.

Contribution

It presents a novel probabilistic active learning approach that autonomously identifies informative measurement locations in neutron scattering experiments using log-Gaussian processes.

Findings

Method reduces measurement time in TAS experiments.

Demonstrated effectiveness on real TAS data.

Outperforms traditional manual search approaches.

Abstract

Neutron scattering experiments at three-axes spectrometers (TAS) investigate magnetic and lattice excitations by measuring intensity distributions to understand the origins of materials properties. The high demand and limited availability of beam time for TAS experiments however raise the natural question whether we can improve their efficiency and make better use of the experimenter's time. In fact, there are a number of scientific problems that require searching for signals, which may be time consuming and inefficient if done manually due to measurements in uninformative regions. Here, we describe a probabilistic active learning approach that not only runs autonomously, i.e., without human interference, but can also directly provide locations for informative measurements in a mathematically sound and methodologically robust way by exploiting log-Gaussian processes. Ultimately, the resulting benefits can be demonstrated on a real TAS experiment and a benchmark including numerous different excitations.