CrystalPilot: A machine-learning based software platform for single crystal neutron diffraction experiments

TL;DR

CrystalPilot is a machine learning software that improves the efficiency and accuracy of neutron diffraction experiments by automatically steering them and reducing background noise.

Contribution

CrystalPilot introduces machine learning methods for adaptive neutron diffraction experiments and novel background noise treatment techniques.

Findings

CrystalPilot uses real-time regression models to steer neutron diffraction experiments efficiently.

Two new methods, KD-Tree Estimation and Ring Feature Extraction, effectively reduce background noise in neutron scattering data.

The platform enhances data quality and accelerates experimental workflows in neutron diffraction.

Abstract

We present an in-house software platform CrystalPilot, designed to enable smart, adaptive single crystal neutron diffraction experiments at the TOPAZ beamline at Oak Ridge National Laboratory. It leverages machine learning methods to provide intelligent experiment steering and advanced algorithms for background noise treatment.

Neutron scattering experiments are inherently high-cost; therefore, automatic steering is essential for optimizing efficiency and maximizing scientific output. CrystalPilot employs real-time regression-based prediction models to predict the data quality and combine with automatic decision-making to steer the experiments in single crystal neutron diffraction. This approach minimizes the neutron beam time, ensuring experiments remain efficient.

Background noise significantly hampers accurate analysis of material properties. We present two machine learning-based methods for background treatment: KD-Tree Estimation and Ring Feature Extraction. The former use a KD-Tree based data structure to estimate the statistical information of background in different regions, while the latter uses a non-uniform Gaussian fitting model to extract the most prominent aluminum ring artifacts in neutron scattering data. Both methods have been tested and demonstrated to effectively remove background noise, enhancing data quality and making them well-suited for different neutron scattering applications.

By combining these algorithms within a cohesive software platform, CrystalPilot ensures that neutron diffraction experiments remain adaptive to real-time conditions, thereby accelerating experimental workflows and improving data fidelity. This integration of machine learning and statistical modeling establishes a robust framework for neutron experiment steering, significantly enhancing throughput, data quality, and scientific discovery.