GoniOwl: convolutional neural network-based sample state detection for collision prevention on synchrotron beamlines

Area of Science:

• Synchrotron Science
• Artificial Intelligence in Scientific Instrumentation
• Crystallography

Background:

• Automated sample exchange is crucial for synchrotron beamlines, especially with complex cryogenic systems.
• Collisions during this process can cause significant hardware damage and downtime.

Purpose of the Study:

• To develop and implement an AI-based system for real-time detection of sample pins during automated exchange.
• To enhance machine protection and prevent collisions in synchrotron beamlines.

Main Methods:

• A compact convolutional neural network (CNN) named GoniOwl was trained on over 8700 images from the I23 macromolecular crystallography beamline.
• The model was augmented for robustness against varying illumination, camera shifts, and occlusions.
• A confidence-gating mechanism was integrated for operator confirmation of uncertain predictions.

Main Results:

• GoniOwl achieved >99% accuracy with millisecond-level inference times.
• The CNN model matched or exceeded the performance of legacy histogram methods and human operators (96% accuracy).
• A closed-loop workflow was established for continuous model improvement through automated auditing of divergent cases.

Conclusions:

• GoniOwl provides a reliable software machine protection layer for automated sample exchange at synchrotron beamlines.
• The AI-driven approach enhances safety, reduces downtime, and is transferable to other beamline environments.
• This technology improves operational efficiency and data collection reliability in structural biology.