Two-Dimensional Energy Histograms as Features for Machine Learning to Predict Adsorption in Diverse Nanoporous...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine learning (ML) in chemical science faces challenges due to a lack of physically informed feature representations.
• Accurate prediction and model interpretability are crucial for advancing ML applications in materials science.

Purpose of the Study:

• To develop novel, physically informed feature representations for adsorption systems.
• To improve the accuracy and interpretability of ML models for predicting adsorption capacity in nanoporous materials.

Main Methods:

• Introduced two-dimensional energy histogram (2D-EH) features derived from probe-adsorbent energies and gradients.
• Applied 2D-EH features to predict single-component adsorption capacity in metal-organic frameworks (MOFs) for various molecules and conditions.
• Evaluated the transferability of ML models trained with 2D-EH features to amorphous nanoporous materials.

Main Results:

• Achieved highly accurate ML models for predicting adsorption capacity (R² ∼ 0.94-0.99).
• Demonstrated that 2D-EH features encode both energetic and structural information, enabling ML models to learn adsorption physics.
• Showed successful transferability of MOF-trained models to other nanoporous materials, outperforming existing structural features.

Conclusions:

• Novel 2D-EH features offer a significant advancement for ML in materials science, enhancing prediction accuracy and interpretability.
• These features facilitate the understanding of adsorption phenomena and the design of new nanoporous materials for gas storage and separation.
• The developed ML approach shows promise for accelerating the discovery of advanced materials.