Reinforcement Learning Informed Growth Dynamics in 2D Materials

The utilization of two-dimensional materials despite their extraordinary physical and chemical properties have been limited in applications like foldable electronics, memristors, water treatment owing to poor understanding of their synthesis protocols. This can be attributed to the limited understanding or absence of data on aspects like substrate selection, possible metastable and defective configurations. While Molecular Dynamics (MD) can provide such atomistic details for various time dependent synthesis protocols, it remains limited in the timescales that it can access. Additionally, it is cumbersome or rather impossible to probe all the possible synthesis protocols. To address this issue, we developed a Deep-Q-Networks (DQN) based reinforcement learning (RL) approach combined with an efficient and accurate force-field model to grow Phosphorene (alpha and beta) polymorphs on different orientations of copper substrate. Our RL agent is assigned to vary growth parameters such as temperature, cooling rate, deposition rate, and substrate crystal orientation. In each iteration, the final configuration is evaluated against the desired phase using SOAP fingerprinting and/or graph isomorphism. With this feedback, the RL agent then conducts a series of exploratory or exploitative iterations to determine the optimal growth parameters for achieving the desired conditions. Also, it has an added advantage of retrieving possible metastable configurations obtained while performing the search for a desired phase. Furthermore, this technique holds promise for integration with self-autonomous experimentation tools, accelerating the development and application of present and future 2D materials.