Universal machine-learning algorithm for predicting adsorption performance of organic molecules based on limited data...

Area of Science:

• Environmental Science
• Computational Chemistry
• Materials Science

Background:

• Adsorption is a key method for removing organic molecules from aqueous solutions.
• Previous machine learning (ML) applications using polyparameter linear free energy relationships (pp-LFERs) showed limitations, especially outside their applicability domain.
• There is a need for improved ML models that can handle diverse organic molecules and predict adsorption accurately.

Purpose of the Study:

• To enhance the applicability of ML methods for predicting organic molecule adsorption.
• To develop and evaluate a chemical-structure (CS) based ML approach for adsorption prediction.
• To compare the performance of different CS feature descriptors and ML algorithms.

Main Methods:

• Utilized a chemical-structure (CS) based approach for machine learning model development.
• Employed two CS feature descriptors: 3D-coordination and simplified molecular-input line-entry system (SMILES).
• Built and compared machine learning models using neural networks (NN) and extreme gradient boosting (XGB).

Main Results:

• The CS-based ML models demonstrated superior performance compared to pp-LFERs based models.
• Models accurately predicted adsorption isotherms, including for isomers like chiral molecules, even when trained on achiral data.
• Extreme gradient boosting (XGB) outperformed neural networks (NN) for adsorption isotherm prediction, with 3D-coordination descriptors showing high effectiveness.

Conclusions:

• A chemical-structure based approach significantly improves the accuracy and applicability of ML models for predicting organic molecule adsorption.
• 3D-coordination descriptors effectively capture molecular structural differences relevant to adsorption interactions.
• XGB models utilizing CS features offer a powerful tool for predicting adsorption behavior of diverse organic compounds.