Continuous discovery of novel 2D materials via dual active learning-driven generative models

Area of Science:

• Materials Science
• Artificial Intelligence
• Computational Chemistry

Background:

• Generative AI accelerates materials discovery but is limited by historical data bias, especially in data-scarce fields like 2D materials.
• Existing generative models often produce repetitive outputs, hindering the discovery of genuinely novel materials.

Purpose of the Study:

• To introduce DuALGen, a dual active learning framework designed to mitigate data bias and enhance diversity in generative materials discovery.
• To enable the exploration of uncharted chemical spaces and facilitate the continuous discovery of new materials.

Main Methods:

• DuALGen employs two coupled active learning loops: a generative loop with dynamic sampling for design space exploration and a predictive loop sampling outliers to counter distribution shift.
• The framework integrates generative models with outlier detection for reliable evaluation of novel candidates.

Main Results:

• Application of DuALGen to 2D materials resulted in the discovery of over 10,000 stable and distinct compounds.
• Thousands of high-performance 2D material candidates for electronic applications were identified.

Conclusions:

• DuALGen effectively enriches data diversity and corrects bias in generative models for materials discovery.
• This self-updating workflow provides a practical pathway for the continuous discovery of novel materials, particularly in data-scarce domains.