A machine learning decision criterion for reducing scan time for hyperspectral neutron computed tomography systems.

Area of Science:

• Materials Science
• Neutron Imaging
• Machine Learning

Background:

• Hyperspectral neutron computed tomography (HNCT) is crucial for materials science, enabling detailed crystallographic and elemental analysis.
• Traditional HNCT methods require extensive data acquisition (days) and numerous projections, limiting real-time feedback and efficiency.
• Existing advanced methods like golden ratio scanning offer improvements but lack automated scan termination based on data quality.

Purpose of the Study:

• To introduce the first machine learning-based autonomous HNCT experiment.
• To develop an intelligent criterion for automatically terminating HNCT scans when sufficient data is acquired.
• To reduce overall measurement time while maintaining high-quality reconstructions.

Main Methods:

• Implemented a machine learning criterion for autonomous scan termination in HNCT.
• Utilized a pre-trained deep neural network for a reference-free image quality assessment.
• Combined image quality metrics with reconstruction difference metrics to guide scan termination.
• Employed a golden ratio scanning protocol with model-based reconstruction for real-time feedback.

Main Results:

• The proposed machine learning criterion successfully terminated HNCT scans autonomously.
• Achieved a reduction in measurement time by approximately a factor of five compared to baseline filtered back projection methods.
• Demonstrated high-quality real-time reconstructions from streaming experimental data.

Conclusions:

• The developed machine learning approach enables efficient and autonomous HNCT experiments.
• This method significantly accelerates sample characterization in materials science by reducing scan times.
• Autonomous termination criteria enhance experimental workflow and provide timely user feedback.