Using Reinforcement Learning for Manipulating and Synthesizing Thin Films

Recent progress in reinforcement learning (RL), and in particular the combination of traditional RL methods with deep neural networks for function approximation, has resulted in tremendous advances in gameplay and sequential decision-making tasks. However, the applications of RL to materials simulation has been limited and best, largely due to difficulties stemming from the very large data requirements for traditional RL.<br/> <br/>Here, we explore the application of RL within simulations and experiments, focused on the manipulation of structures within ferroelectric materials. We show that utilizing RL agents, it is possible to learn optimal defect positioning within a lattice-based ferroelectrics simulation environment modeled with nearest neighbor interactions. Investigation of the agent’s policy reveals underlying information about the length scales of interactions between defects in the simulation. We then proceed to a more complex molecular dynamics environment of the synthesis of Tellurene, and proceed to run thousands of simulations of deposition of Te to assemble into Tellurene, and then employ off-policy RL to learn appropriate actions to maximize the crystallinity and target phase of the resultant structure. Finally, results on manipulation of materials experimentally with RL implemented on a microscope are discussed, with a view towards the need for better integrating simulations in such design and characterization workflows. This work was supported by Center for Nanophase Materials Sciences (CNMS), which is a US Department of Energy, Office of Science User Facility at Oak Ridge National Laboratory.