Predicting Lewis Acidity: Machine Learning the Fluoride Ion Affinity of p-Block-Atom-Based Molecules

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Lewis Acid Characterization

Background:

• Lewis acid strength is crucial in chemistry, often quantified by fluoride ion affinity (FIA).
• Traditional FIA calculations using quantum chemistry are computationally intensive.
• A need exists for faster, accurate methods to assess Lewis acid strength.

Purpose of the Study:

• To develop a machine learning model for predicting fluoride ion affinity (FIA).
• To create a comprehensive dataset of FIA values for training machine learning models.
• To enable rapid and accurate assessment of Lewis acid strength.

Main Methods:

• Calculated 48,986 fluoride ion affinity (FIA) values using high-level quantum chemistry (RI-DSD-BLYP-D3(BJ)/def2-QZVPP//PBEh-3c).
• Developed two message-passing graph neural networks (FIA-GNN) trained on the FIA49k dataset.
• Input for the model is the SMILES string of the Lewis acid.

Main Results:

• The FIA-GNN model predicts gas and solution phase FIA values with a mean absolute error of 14 kJ/mol (r²=0.93).
• The model achieves high accuracy across a wide energetic range (750 kJ/mol).
• Model performance was validated through case studies, including predictions for Cambridge Structural Database molecules and literature catalysis data.

Conclusions:

• FIA-GNN provides a computationally efficient and accurate method for predicting Lewis acid strength.
• The FIA49k dataset and FIA-GNN model are valuable resources for computational and synthetic chemists.
• The study demonstrates the power of graph neural networks in chemical property prediction.