The Black Hole Strategy: Gravity-Based Representative Sampling for Frugal Graph Learning on Metal-Organic Framework...

Area of Science:

• Materials Informatics
• Machine Learning
• Data Science

Background:

• Large materials databases benefit from graph-based representations for machine learning.
• Dense datasets can lead to redundancy, noise, and increased computational costs without performance gains.

Purpose of the Study:

• To introduce a novel representative sampling method, the Black Hole Strategy, for creating compact and informative subsets from large materials datasets.
• To demonstrate the efficacy of this strategy in preserving essential structural and property diversity for machine learning applications.

Main Methods:

• Developed a gravity-based representative sampling method called the Black Hole Strategy.
• Applied the strategy to sparsify datasets of metal-organic frameworks (MOFs).
• Trained and evaluated graph neural networks (GraphSAGE, GCN, GAT) on both full and sparsified datasets.

Main Results:

• Machine learning models trained on Black Hole-sparsified datasets achieved comparable or superior performance to full-dataset models.
• Significant reductions in data points, memory usage, and training time were observed.
• Critical structure-property relationships, like pore-limiting diameter, were preserved under substantial data sparsification.

Conclusions:

• The Black Hole Strategy is a principled, frugal, and robust approach for machine learning in materials science.
• It enables efficient, interpretable, and scalable discovery workflows.
• The method advances FAIR data practices by enhancing data management, reusability, and interoperability.