PoroNet: An Intrinsically Interpretable Pore Graph Neural Network for Resolving Pore-Level Adsorption in...

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Machine learning (ML) models are used to predict adsorption in metal-organic frameworks (MOFs) for gas storage and separations.
• The 'black box' nature of traditional ML models hinders the design of novel MOFs.
• Interpretable ML is needed to bridge the gap between prediction and material design.

Purpose of the Study:

• Introduce PoroNet, an interpretable graph neural network (GNN) for MOF adsorption prediction.
• Enable accurate prediction of gas uptake and deliverable capacity in MOFs.
• Provide pore-level insights into adsorption behavior for MOF design.

Main Methods:

• Developed PoroNet, a GNN based on a pore graph representation of MOFs.
• Trained PoroNet for accurate prediction of hydrogen (H2) uptake and alkane adsorption.
• Utilized direct supervised learning and latent representations for pore-level contribution analysis.

Main Results:

• PoroNet achieved high accuracy in predicting H2 uptake and deliverable capacity in MOFs.
• The model accurately predicted adsorption on a benchmark dataset with various alkane adsorbates.
• PoroNet effectively learned pore-level contributions to total adsorption, enhancing interpretability.
• Demonstrated data efficiency, requiring fewer training runs compared to standard approaches.

Conclusions:

• PoroNet provides interpretable predictions of MOF adsorption, facilitating MOF design and discovery.
• The model identifies key pore properties governing adsorption and guides pore engineering.
• PoroNet is a valuable tool for high-throughput screening and developing MOF design rules for cryogenic H2 storage.
• Highlights the potential of interpretable ML in accelerating scientific and material discovery.