Utilizing Machine Learning Models for Predicting Diamagnetic Susceptibility of Organic Compounds

Area of Science:

• Computational Chemistry
• Organic Chemistry
• Materials Science

Background:

• Magnetic susceptibility is crucial for understanding molecular structure and reactions.
• Traditional methods for magnetic susceptibility calculation are inefficient and resource-intensive.

Purpose of the Study:

• To develop a novel, efficient method for predicting molar magnetic susceptibility of organic molecules.
• To establish quantitative structure-property relationships (QSPR) for magnetic susceptibility.

Main Methods:

• Utilized a dataset of 382 organic molecules with known molar magnetic susceptibility.
• Employed six molecular fingerprinting techniques (RDKit, Morgan, MACCS, atom pair, Avalon, topology) as input features.
• Trained seven machine learning models (Random Forest, AdaBoost, Gradient Boosting, Extra Trees, Elastic Net, SVM, MLP).

Main Results:

• The combination of atom pair fingerprint and Multilayer Perceptron (MLP) model achieved high predictive accuracy.
• Achieved R-squared values of 0.88 (validation set) and 0.90 (test set).
• Demonstrated exceptional predictive performance for molar magnetic susceptibility.

Conclusions:

• The developed QSPR-MLP model offers a highly accurate and efficient approach for predicting organic molecule magnetic properties.
• This method significantly accelerates research and development by reducing experimental and computational costs.
• Represents a breakthrough in rapid property screening for chemical compounds.