Deep Reinforcement Learning for Dopants Design

Mentors: Ray Lei, Weike Ye, Joseph Montoya

Project summary Introducing low concentrations of impurities (i.e., dopants) to crystal materials via rational design can potentially improve their performance drastically. Current in-silico dopant design approaches suffer from imbalanced exploration-exploitation, poor scalability, inability to generalize, among other difficulties. This project proposes to use reinforcement learning (RL) to tackle the dopants design problem. We first propose a framework for casting generic dopantx design problemx into RL games. Then we designed an invariant and explainable policy network and use policy gradient approach to learn optimal doping policy. We demonstrate unique advantages of this policy-based design approach, including high search efficiency, ability to transfer to larger systems and systems with different dopants concentrations. This project leads to lots of exciting future research directions.

Project outcome

• Utility patent submission (currently under examination)
• Python package for RL-based dopant design, now utilized by the Energy & Environment team
• Presented the results to the experimental team to facilitate decision-making