Predicting Hydrocarbon Strain Energy via a Group Equivalent Machine Learning Approach.

Area of Science:

• Computational chemistry
• Organic chemistry
• Machine learning

Background:

• Strain energy quantifies molecular steric and configurational properties.
• Estimating strain energy via quantum chemistry requires initial nuclear coordinates, often unknown in molecular design.
• Predicting strain energy is crucial for screening and generating novel molecular candidates.

Purpose of the Study:

• To develop a machine learning model for predicting hydrocarbon strain energies.
• To utilize Benson group equivalents and molecular fingerprints for accurate strain energy estimation.
• To provide a computational tool for molecular design that bypasses the need for initial structural data.

Main Methods:

• A machine learning approach was developed using Benson group equivalents.
• A featurization strategy combined group equivalent counts with molecular fingerprints.
• Data were derived from electronic structure calculations on 166 synthesized strained hydrocarbons.

Main Results:

• The study evaluated the predictive accuracy of various statistical learning methods for strain energy.
• The developed model demonstrated effective prediction of hydrocarbon strain energies.
• The performance merits and limitations of different machine learning models were discussed.

Conclusions:

• Machine learning, utilizing Benson group equivalents and molecular fingerprints, offers an effective method for predicting hydrocarbon strain energies.
• This approach facilitates molecular design by enabling strain energy estimation without requiring initial molecular coordinates.
• The findings contribute to computational chemistry by providing a practical tool for assessing molecular properties.