Learning from virtual experiments to assist users of Small Angle Neutron Scattering in model selection

Area of Science:

• Materials Science
• Neutron Scattering Physics
• Computational Science

Background:

• Small Angle Neutron Scattering (SANS) is crucial for characterizing nanoscale structures.
• Model selection in SANS data analysis can be complex and time-consuming.
• Advancements in machine learning offer new approaches to complex scientific data interpretation.

Purpose of the Study:

• To develop an automated tool for SANS model selection using machine learning.
• To assist SANS users in interpreting experimental data more efficiently.
• To leverage virtual SANS experiments for training predictive models.

Main Methods:

• Generation of a large dataset (~260,000) of virtual SANS experiments using Monte Carlo simulations.
• Training an ensemble of Convolutional Neural Networks (CNNs) on 2D scattering patterns.
• Utilizing computer vision techniques for pattern recognition and classification.

Main Results:

• The CNN-based recommendation system achieved high accuracy in classifying 46 different SANS models.
• The system demonstrated the ability to learn the relationship between scattering patterns and underlying models.
• Successful testing of the trained network with real experimental SANS data.

Conclusions:

• The developed recommendation system shows significant potential for improving SANS data analysis workflows.
• Virtual SANS experiments are valuable for creating training datasets for machine learning models.
• This approach can streamline the model selection process for SANS users.