Density-aware active learning for materials discovery: a case study on functionalized nanoporous materials.

Area of Science:

• Chemistry and Materials Science
• Machine Learning
• Data Science

Background:

• High-performance machine learning requires large datasets, which are costly and time-consuming to acquire in fields like chemistry and materials science.
• Active learning (AL) methods reduce data needs by iteratively selecting informative samples, but regression tasks present unique challenges due to uncertainty estimation and continuous output spaces.

Purpose of the Study:

• To develop an efficient active learning strategy for regression tasks in large design spaces.
• To minimize training dataset size while maintaining predictive accuracy for machine learning models.

Main Methods:

• Introduced density-aware greedy sampling (DAGS), an active learning algorithm for regression.
• Integrated uncertainty estimation with data density to guide sample selection.
• Evaluated DAGS on synthetic and real-world datasets of nanoporous materials (MOFs, COFs) for separation applications.

Main Results:

• DAGS consistently outperformed random sampling and existing state-of-the-art active learning methods.
• Achieved effective training of regression models with significantly reduced data points.
• Demonstrated strong performance even on high-dimensional datasets.

Conclusions:

• Density-aware greedy sampling (DAGS) is a highly effective active learning method for regression tasks.
• DAGS offers a robust solution for reducing data acquisition costs in materials science and chemistry.
• The method shows promise for accelerating the discovery and optimization of functional materials.