High-efficient ab initio Bayesian active learning method and applications in prediction of two-dimensional functional...

Area of Science:

• Computational Materials Science
• Machine Learning in Materials Discovery
• Materials Informatics

Background:

• Traditional materials discovery relies on inefficient trial-and-error methods.
• Machine learning accelerates discovery but faces challenges like limited data and unbalanced property distributions.
• Predicting functional materials with specific properties requires efficient screening methods.

Purpose of the Study:

• To develop and validate an ab initio Bayesian active learning method.
• To accelerate the prediction of functional materials with high efficiency and accuracy.
• To address challenges of limited data and unbalanced property distributions in materials science.

Main Methods:

• Combined active learning with high-throughput ab initio calculations.
• Applied the method to a dataset of 3119 two-dimensional hexagonal binary compounds.
• Screened for materials with maximal electric polarization and appropriate photovoltaic band gaps.

Main Results:

• Successfully identified materials with desired properties, including maximal electric polarization and suitable photovoltaic band gaps.
• Significantly reduced computational costs by calculating only a fraction of potential candidates compared to random search.
• Demonstrated high efficiency and accuracy in screening materials with unbalanced property distributions.

Conclusions:

• The ab initio Bayesian active learning method offers significant advantages for materials discovery, especially with unbalanced property distributions.
• This approach drastically reduces computational expenses while maintaining high accuracy.
• The method is readily applicable to discovering a wide range of advanced functional materials.