Towards Augmented Microscopy with Reinforcement Learning-Enhanced Workflows.

Area of Science:

• Materials Science
• Microscopy
• Artificial Intelligence

Background:

• Automating complex microscopy workflows is crucial for efficiency.
• Scanning transmission electron microscopy (STEM) requires precise beam alignment.
• Current alignment methods can be time-consuming and require expert knowledge.

Purpose of the Study:

• To implement and evaluate reinforcement learning (RL) for automating STEM operations.
• To develop an RL agent for autonomous electron beam alignment without prior knowledge.
• To assess the impact of virtual environment design on RL performance.

Main Methods:

• Designed a virtual, prototypical RL environment for STEM beam alignment.
• Developed and trained a neural network using RL within the simulator.
• Evaluated alignment accuracy and convergence across various hyperparameters.
• Deployed the trained RL model on an actual STEM microscope.

Main Results:

• Demonstrated robust convergence of the RL model across a wide hyperparameter space in simulation.
• The on-microscope RL model achieved goal alignment with minimal training.
• Validated the effectiveness of the virtual environment design for real-world deployment.
• Confirmed that RL can automate beam alignment without extensive algorithm design.

Conclusions:

• Reinforcement learning offers a viable method for automating STEM operations.
• Virtual environment simulation is key to successful RL implementation in microscopy.
• This approach advances the integration of machine learning in electron microscopy.