Data-driven simulations for training AI-based segmentation of neutron images

Area of Science:

• Materials science
• Neutron imaging
• Artificial intelligence

Background:

• Neutron interferometry enables nanoscale material characterization (1 nm–10 µm).
• Limited experimental time restricts the number of images for AI-driven analysis.
• Supervised AI models require extensive annotated data, often unavailable due to time constraints.

Purpose of the Study:

• To develop a data-driven simulation framework for generating synthetic neutron images.
• To augment limited experimental datasets for improved AI model training.
• To reduce manual labor and increase confidence in AI-based image segmentation accuracy.

Main Methods:

• Utilized a data-driven simulation framework incorporating Johnson family probability density functions (PDFs).
• Implemented a multi-step process: PDF estimation, validation, mask design, intensity generation, and AI model training.
• Applied the framework to segment four-dimensional images of nine calibration phantoms using AI models.

Main Results:

• Successfully generated synthetic images to supplement measured data for AI training.
• AI models trained with combined synthetic and measured data achieved accurate segmentation.
• Demonstrated the framework's effectiveness in overcoming data scarcity for neutron imaging analysis.

Conclusions:

• The developed simulation framework effectively addresses data limitations in neutron interferometry.
• Synthetic data generation enhances AI model performance for material segmentation.
• This approach offers a viable solution for automated analysis in neutron imaging experiments.