Learning with Delayed Rewards-A Case Study on Inverse Defect Design in 2D Materials

Area of Science:

• Materials Science
• Computational Materials Science
• Artificial Intelligence in Materials

Background:

• Defect dynamics are crucial for material functionality in technologies like catalysis and microelectronics.
• Understanding and designing defect arrangements is challenging due to intermediate states and energy barriers.
• Current inverse design methods struggle with these complexities.

Purpose of the Study:

• To introduce a novel reinforcement learning (RL) strategy for efficient defect configurational space search.
• To overcome energy barriers in defect transformation for inverse defect design.
• To identify optimal defect arrangements for targeted material functionality.

Main Methods:

• Utilized Monte Carlo Tree Search (MCTS) with delayed rewards, a type of RL.
• Applied the method to a representative case of two-dimensional (2D) molybdenum disulfide (MoS2) with sulfur (S) vacancies.
• Compared the MCTS approach with genetic algorithms for global optimization.

Main Results:

• The RL strategy efficiently sampled the defect configurational space and overcame energy barriers for S vacancies in MoS2 (1.5-8% concentration).
• Observed evolution from random S vacancies to extended S line defects, consistent with experimental findings.
• MCTS with delayed rewards required fewer evaluations and yielded higher quality solutions than genetic algorithms.

Conclusions:

• The developed RL strategy accelerates the inverse design of defects in materials.
• Identified optimal pathways for defect transformation and arrangement in 2D MoS2.
• Discussed implications for 2H to 1T phase transitions in MoS2, impacting material properties.