RetNeXt: A Pretrained Model for Transfer Learning Across the MOF Adsorption Space.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Metal-organic frameworks (MOFs) possess high porosity and tunable chemistry, making them promising for gas adsorption.
• The vast chemical space of MOFs challenges traditional exploration methods for identifying optimal materials.
• Current machine learning (ML) models for MOF adsorption properties are often single-task, requiring retraining from scratch for new properties and data.

Purpose of the Study:

• To develop a more data-efficient and generalizable machine learning framework for predicting MOF adsorption properties.
• To leverage multitask and transfer learning to overcome data scarcity in training predictive models.
• To create a foundation model (RetNeXt) applicable to diverse adsorption tasks and conditions.

Main Methods:

• Extended a previous framework combining potential energy surfaces with convolutional neural networks.
• Introduced multitask and transfer learning to enhance model generalization.
• Developed RetNeXt, a multitask pretrained model trained on 3.2 million adsorption-related data points.

Main Results:

• RetNeXt demonstrates superior performance compared to conventional single-task transfer learning approaches.
• Achieved up to a 100-fold increase in sample efficiency compared to training models from scratch.
• The model shows effective generalization across different gases and conditions, even with limited data.

Conclusions:

• RetNeXt provides a robust and data-efficient foundation for machine learning-based adsorption modeling of MOFs.
• The multitask pretrained model facilitates rapid adaptation to new adsorption prediction tasks and domains.
• This approach accelerates the discovery and design of MOFs for gas adsorption applications.