Reinforcement learning for optimizing magnetic skyrmion creation.

Area of Science:

• Spintronics
• Artificial Intelligence
• Condensed Matter Physics

Background:

• Topologically stabilized skyrmions are significant spin structures with potential in data storage and computing.
• Traditional skyrmion generation methods rely on complex, trial-and-error field or current tuning.
• These methods often result in numerous intermediate phases, hindering efficient generation.

Purpose of the Study:

• To propose and evaluate a phase-control method using reinforcement learning (RL) for optimizing skyrmion generation.
• To develop an AI-driven approach that automates and enhances the reliability of skyrmion synthesis.
• To reduce the complexity and manual intervention required in generating spintronic structures.

Main Methods:

• Implemented a reinforcement learning framework with a reward system based on topological number and feature states.
• Trained the RL network using physical insights to guide the optimization of field-tuning sequences.
• Validated the network's ability to autonomously generate skyrmions after training.

Main Results:

• The RL network progressively learned and optimized field sequences for skyrmion generation.
• The trained network demonstrated autonomous and reliable generation of skyrmions.
• The method significantly reduced the need for manual adjustments and trial-and-error processes.

Conclusions:

• Reinforcement learning offers an effective strategy for optimizing field control in skyrmion generation.
• This AI-driven approach advances spintronic simulations and experimental implementations.
• The method has potential for generating other spintronic structures like domain walls and vortices.