Reinforcement Learning-Assisted Ferroelectric Domain Wall Design Using a Machine Learning Phase-Field Surrogate

Precise manipulation of ferroelectric domain walls (DWs) has garnered increasing interest for applications in DW memory devices. Although recent advancements in scanning probe microscopy-based automated experiments have improved tip control efficiency, achieving real-time optimization of the tip trajectory to configure arbitrary domain structures remains challenging. In this study, we introduce a reinforcement learning (RL) framework for autonomous DW manipulation, leveraging a 3-D machine learning phase-field surrogate model to accelerate environment dynamics. The RL agent is trained to optimize the tip trajectory, enabling it to achieve target DW configurations within both single-goal and goal-augmented RL frameworks while effectively generalizing across a range of target structures. This framework offers a promising approach for facilitating the real-time design of 2-D ferroelectric DWs.