Generalizable and Transferable Machine Learning Enables Accelerated Metal-Organic Framework Discovery in Gas...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Gas separation is critical for industrial processes, climate mitigation, and clean energy.
• Metal-organic frameworks (MOFs) show promise for adsorption-based separations due to their tunability.
• Identifying optimal MOFs is difficult due to vast structural diversity and high simulation costs.

Purpose of the Study:

• To develop a generalizable machine learning framework for accelerated discovery of MOFs for gas separation.
• To create a curated database (BiMix-Bench) of MOFs and gas mixtures for training and validation.
• To enable data-efficient adaptation of models for new separation tasks.

Main Methods:

• Curated a database (BiMix-Bench) of ~125,900 MOFs and five binary gas mixtures.
• Developed LightGBM regressor (LGBMR) models for predicting gas uptakes and selectivity.
• Evaluated zero-shot and few-shot transfer learning performance using CO2/H2 as a case study.

Main Results:

• LGBMR models achieved high predictive accuracy for gas uptakes (R² = 0.93, 0.92) and selectivity (R² = 0.95).
• Zero-shot predictions showed limited out-of-distribution accuracy.
• Transfer learning with a small number of simulations (N=204) enabled efficient adaptation and identification of top-performing MOFs.

Conclusions:

• The developed framework enables scalable, data-driven discovery of advanced adsorbents for gas separation.
• The approach facilitates rapid identification and validation of MOFs through data-efficient adaptation.
• This interpretable framework addresses the challenges of MOF screening for complex separation tasks.