Adsorption Isotherm Predictions for Multiple Molecules in MOFs Using the Same Deep Learning Model

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Metal-organic frameworks (MOFs) offer tunable adsorption properties for energy and environmental applications.
• Molecular simulation screens MOFs but can be time-consuming.
• Need for faster screening methods like machine learning (ML) is critical.

Purpose of the Study:

• To develop a machine learning model capable of predicting full adsorption isotherms for molecules in MOFs.
• To demonstrate a novel training approach using "alchemical" species to predict real adsorbate behavior.
• To accelerate the discovery and application of MOFs.

Main Methods:

• Trained a multilayer perceptron (MLP) neural network.
• Used "alchemical" species (defined by force-field parameters) for training.
• Represented MOFs using simple descriptors (geometric, chemical).

Main Results:

• The MLP accurately predicted adsorption isotherms for six real molecules (Ar, Kr, Xe, CH4, C2H6, N2) in various MOFs.
• The model successfully predicted adsorption in both hypothetical and existing MOFs.
• Demonstrated the feasibility of predicting adsorption for molecules not included in the training set.

Conclusions:

• This study presents the first ML model predicting full adsorption isotherms using "alchemical" training species.
• The results validate a new training philosophy for ML models in MOF adsorption.
• This approach can be expanded to predict adsorption across diverse MOFs, adsorbates, and conditions.