| Abstract Scope |
Reinforcement learning (RL) has become increasingly used within computational materials science, however, the use of RL within core experimental realms is very limited. Here, I will review our work on utilizing RL in both simulations and experiments. We begin with a focus on the algorithmic developments, including load-balanced RL algorithms capable of running multiple agents to solve RL problems more efficiently, and their implementation on understanding and manipulating atomic scale structures in 2D materials and ferroelectrics, in simulated environments.
Next, we will focus on more recent implementation of RL workflows within experimental setups for manipulating ferroelectric domain walls. Automated experiments capture data relevant to update a physics-informed dynamics model for state transitions, and RL agents can be trained on this model before being deployed on the instrument for autonomous manipulation of domain walls. Finally, we discuss the use of curiosity-driven RL towards improving the theory-experiment feedback loop in real-time. |