Predicting Solvent-Dependent Nucleophilicity Parameter with a Causal Structure Property Relationship.

Area of Science:

• Synthetic chemistry
• Computational chemistry
• Machine learning applications

Background:

• Solvent-dependent reactivity is crucial for controlling reaction selectivity in chemical synthesis.
• There is a growing need for predictive models of reactivity in various solvents, especially sustainable ones.

Purpose of the Study:

• To develop accurate machine learning models for predicting the nucleophilicity parameter (N).
• To investigate the influence of solvent properties and nucleophile descriptors on reactivity predictions.
• To explore the efficiency of different computational methods.

Main Methods:

• Utilized a Causal Structure Property Relationship (CSPR) approach.
• Employed machine learning (Extra Trees algorithm) for prediction.
• Represented nucleophiles with electronic and steric descriptors, and solvents using PCA descriptors from the ACS Solvent Tool.
• Compared DFT descriptors with faster PM6 descriptors.

Main Results:

• Achieved excellent prediction of the nucleophilicity parameter (N) with R² = 0.93.
• 81.6% of predictions were within ±2.0 of experimental values.
• Identified solvent-dependent HOMO energy and Hirshfeld charge as key descriptors.
• PM6 descriptors reduced computation time by 8.7-fold with a minor decrease in prediction accuracy.

Conclusions:

• Machine learning models can reliably predict nucleophilicity across different solvents.
• CSPR provides insights into the physicochemical relationships governing reactivity.
• The study supports the use of efficient computational methods for reactivity prediction in sustainable chemistry.