Direct prediction of gas adsorption via spatial atom interaction learning

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Physisorption in crystalline porous materials is key for sustainable separation, greenhouse gas capture, and energy storage.
• Current methods for screening these materials are slow and lack precision due to the absence of advanced predictive models.

Purpose of the Study:

• To develop an end-to-end deep learning model for accurate and rapid prediction of adsorption in crystalline porous materials.
• To enable direct structure-adsorption prediction using only atomic coordinates and element types.

Main Methods:

• Introduction of DeepSorption, a spatial atom interaction learning network.
• Development of the Matformer module to capture global structure and local atomic interactions.
• Utilizing atomic coordinates and chemical element types as direct input for prediction.

Main Results:

• DeepSorption achieves accurate and fast direct structure-adsorption prediction.
• The model demonstrates a 20-35% reduction in mean absolute error compared to existing methods like CGCNN and descriptor-based machine learning.
• Prediction accuracy surpasses Grand Canonical Monte Carlo simulations.

Conclusions:

• DeepSorption offers a universal framework for predicting physicochemical properties of crystalline materials.
• The model's foundation in interatomic interactions ensures broad applicability and understanding.
• Accelerates the discovery and design of novel porous materials for critical applications.