Gaussian representation for image recognition and reinforcement learning of atomistic structure

Area of Science:

• Computational chemical physics
• Materials informatics
• Machine learning applications

Background:

• Machine learning (ML) in computational chemical physics relies heavily on atomistic structure representation for efficiency.
• Effective representations are crucial for accelerating structure searches and improving property predictions.

Purpose of the Study:

• To investigate the impact of different image representations on a reinforcement learning algorithm's ability to identify atomistic structures.
• To compare the performance of radial Gaussian broadening and angular information representations against a baseline one-hot encoding.

Main Methods:

• Utilized the Atomistic Structure Learning Algorithm (ASLA), a reinforcement learning approach.
• Tested ASLA on two planar atomistic structures: ideal graphene and graphene with a grain boundary.
• Evaluated various image representations, including one-hot encoding, radial Gaussian broadening, and angular information inspired by the smooth overlap of atomic positions (SOAP) method.

Main Results:

• Radial Gaussian broadening of atomic positions significantly benefited the reinforcement learning process for structure identification.
• ASLA demonstrated the capability to optimize the broadening hyperparameters of the Gaussians during the structural search.
• Incorporating angular information did not lead to further speedups in ASLA's performance.

Conclusions:

• The choice of image representation critically influences the success of ML algorithms in materials science.
• Radial Gaussian broadening emerges as a beneficial representation for reinforcement learning-based structure discovery.
• Further research may explore hybrid representations to potentially enhance performance beyond radial broadening alone.