Learning in continuous action space for developing high dimensional potential energy models.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Reinforcement learning (RL) excels in discrete action spaces like games.
• Materials discovery often involves complex, continuous action spaces, posing significant computational challenges.
• Existing optimization methods struggle with scalability and convergence in high-dimensional continuous domains.

Purpose of the Study:

• To develop an efficient and scalable reinforcement learning strategy for continuous action spaces.
• To address limitations of traditional and gradient-based methods in materials design and discovery.
• To enable effective exploration and exploitation in high-dimensional potential energy landscapes.

Main Methods:

• Introduced a novel RL strategy based on decision trees.
• Incorporated modified rewards for improved exploration.
• Implemented efficient sampling during playouts and a "window scaling scheme" for enhanced exploitation.

Main Results:

• Successfully benchmarked the approach against global optimization schemes and state-of-the-art policy gradient methods.
• Demonstrated efficacy in parameterizing potential models for 54 elemental systems and alloys.
• Analyzed error trends, linking them to elemental structural diversity and energy surface smoothness.

Conclusions:

• The developed RL strategy enables efficient and scalable search in continuous action spaces.
• This approach is broadly applicable to physical science problems, including materials discovery and design.
• The method offers a powerful alternative to existing optimization techniques for complex scientific challenges.