Estimating the Hydrogen Bond Strength by Machine Learning Approaches.

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Accurate prediction of hydrogen bond energy is crucial for understanding molecular interactions.
• Existing methods for calculating hydrogen bond energy can be computationally expensive.

Purpose of the Study:

• To investigate machine learning approaches for predicting hydrogen bond energy.
• To identify key molecular descriptors that influence hydrogen bond energy.

Main Methods:

• Utilized support vector regression combined with gradient boosting.
• Employed Löwdin partial charges and bond orders from BLYP or B3LYP with the def2-SVP basis set.
• Explored the semiempirical GFN2-xTB approach for Mulliken partial charges and Wiberg bond orders.

Main Results:

• Achieved a mean absolute percentage error of 3% with the best models.
• Demonstrated significant improvement over previous predictive models.
• GFN2-xTB approach yielded a 4% mean absolute percentage error.

Conclusions:

• Machine learning models can effectively predict hydrogen bond energy.
• Specific descriptors like Löwdin partial charges and BLYP/B3LYP bond orders provide high accuracy.
• The GFN2-xTB approach offers a viable alternative for feature generation.