Informative Training Data for Efficient Property Prediction in Metal-Organic Frameworks by Active Learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Data-driven material discovery often struggles with expensive-to-compute or measure target properties.
• Efficiently selecting informative samples for training is crucial in low-data regimes.
• Metal-organic frameworks (MOFs) offer vast combinatorial possibilities but require effective property prediction.

Purpose of the Study:

• To develop a novel active learning algorithm for constructing diverse and informative training sets in data-driven material discovery.
• To apply this algorithm to predict band gap and adsorption properties of MOFs.
• To enhance the efficiency and accuracy of materials property prediction, especially with limited or imbalanced data.

Main Methods:

• A regression tree-based active learning algorithm was developed.
• Low-dimensional descriptors based on stoichiometric and geometric properties were used for MOF feature space representation.
• The algorithm leverages regression tree partitions to select new samples for training set expansion.

Main Results:

• The proposed method constructs smaller training datasets for regression models compared to existing active learning approaches.
• It achieves more efficient prediction of MOF properties (band gap, adsorption) with lower variance.
• The algorithm demonstrates significant benefits with unevenly distributed and imbalanced label data.

Conclusions:

• The regression tree-based active learning algorithm effectively accelerates materials discovery by improving structure-property relationship analysis.
• It provides a unique tool for efficient analysis of complex materials data, especially in low-data scenarios.
• The method enhances the quality of training sets, leading to more efficient and accurate predictive models for novel materials like MOFs.