Adaptive X-ray imaging with reinforcement learning

Area of Science:

• Materials Science
• Biological Imaging
• Environmental Science

Background:

• Synchrotron light sources are crucial for high-intensity X-ray imaging but are limited and in high demand.
• Standard raster scanning methods are inefficient for sparse samples, wasting time on uninformative areas.

Purpose of the Study:

• To develop a more efficient X-ray imaging technique by adaptively distributing exposure.
• To maximize information gain within a limited time budget for X-ray microscopy.

Main Methods:

• Formulated adaptive X-ray scanning as a reinforcement learning problem.
• Developed agents to generate sequential exposure maps based on previous measurements.
• Simulated the adaptive illumination strategy.

Main Results:

• Simulations showed adaptive illumination can accelerate X-ray measurements by up to an order of magnitude.
• The approach intelligently directs exposure to informative regions, reducing scan time.
• Successfully deployed trained agents on an X-ray fluorescence beamline.

Conclusions:

• Reinforcement learning offers a powerful framework for optimizing X-ray imaging acquisition.
• Adaptive illumination significantly enhances the efficiency of synchrotron-based X-ray microscopy.
• This method holds promise for accelerating research in various scientific fields.