Gas adsorption meets geometric deep learning: points, set and match.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Metal-organic frameworks (MOFs) possess high porosity and surface area, making them suitable for gas adsorption.
• The vast chemical space of MOFs hinders exploration via traditional methods.
• Machine learning (ML) pipelines, typically descriptor-based, struggle to fully utilize 3D structural data.

Purpose of the Study:

• To introduce AIdsorb, a novel descriptor-free framework for predicting gas adsorption properties.
• To demonstrate AIdsorb's ability to directly process raw 3D structural information.
• To evaluate AIdsorb's performance against conventional descriptor-based ML pipelines.

Main Methods:

• The AIdsorb framework treats material structures as point clouds.
• A deep learning algorithm designed for point cloud analysis is employed.
• The framework was applied to predict methane (CH4) uptake in MOFs and carbon dioxide (CO2) uptake in covalent organic frameworks (COFs).

Main Results:

• AIdsorb achieved superior performance in predicting methane uptake in MOFs compared to a conventional descriptor-based pipeline.
• The framework demonstrated transferability by accurately predicting carbon dioxide uptake in COFs.
• The descriptor-free approach effectively utilizes raw structural information.

Conclusions:

• AIdsorb offers a powerful and efficient method for screening materials for gas adsorption.
• The framework's ability to process raw structural data overcomes limitations of descriptor-based methods.
• AIdsorb's applicability extends beyond MOFs and COFs to broader materials science applications.