Exploring Nanocluster Potential Energy Surfaces via Deep Reinforcement Learning: Strategies for Global Minimum...

Area of Science:

• Computational Materials Science
• Nanomaterials Discovery
• Artificial Intelligence in Chemistry

Background:

• Identifying global minimum (GM) configurations in nanoclusters is challenging due to complex potential energy landscapes with many local minima.
• Traditional methods like genetic algorithms and basin hopping struggle with convergence and navigating large, complex energy surfaces, especially for diverse and large nanoclusters.

Purpose of the Study:

• To develop and validate a novel deep reinforcement learning (DRL) framework for efficient exploration of nanocluster potential energy surfaces (PES).
• To identify global minimum (GM) configurations and other low-energy states in various mono- and multimetallic nanoclusters.

Main Methods:

• Development of a specialized deep reinforcement learning (DRL) framework tailored for potential energy surface (PES) exploration.
• Application of the DRL framework to diverse nanocluster systems, including mono- and multimetallic compositions.
• Evaluation of the framework's performance with increasing cluster size and feature vector dimensions.

Main Results:

• The DRL framework successfully navigated complex energy landscapes to identify GM configurations and low-energy states in nanoclusters.
• The model demonstrated remarkable adaptability and sustained efficiency across various nanocluster types and sizes.
• The framework's performance remained robust even with increasing feature vector dimensions.

Conclusions:

• Deep reinforcement learning (DRL) offers a powerful and adaptable methodology for efficient global minimum searches in nanoclusters.
• This novel DRL framework represents a significant advancement for nanomaterials discovery and optimization in materials science.
• The approach shows considerable potential for accelerating the identification of novel functional nanomaterials.