Accurate Thermochemistry with Small Data Sets: A Bond Additivity Correction and Transfer Learning Approach

Area of Science:

• Computational chemistry
• Machine learning
• Chemical informatics

Background:

• Accurate prediction of molecular properties, especially thermochemistry, is crucial for computational simulations like reaction mechanism generation.
• Traditional machine learning models often require extensive datasets, which are frequently unavailable for high-accuracy quantum mechanical or experimental data.
• Existing methods for estimating thermochemical parameters can be limited in scope and accuracy.

Purpose of the Study:

• To develop accurate and rapid machine learning models for predicting molecular thermochemistry.
• To address the challenge of limited high-accuracy training data by employing transfer learning and bond additivity corrections.
• To create versatile thermochemistry predictors applicable to organic compounds with heteroatoms.

Main Methods:

• Calculated new high-level datasets and derived bond additivity corrections to enhance enthalpy of formation predictions.
• Employed a transfer learning technique to train neural network models, enabling good performance with smaller high-accuracy datasets.
• Selected training data for entropy models to effectively capture conformational effects.

Main Results:

• Developed neural network models that predict thermochemical properties with accuracy approaching experimental and coupled cluster levels.
• Achieved good model performance despite using relatively small sets of high-accuracy data through transfer learning.
• Created generally applicable thermochemistry predictors for organic compounds containing oxygen and nitrogen heteroatoms.

Conclusions:

• The developed machine learning models offer a versatile and accurate alternative to conventional methods for estimating thermochemical parameters.
• Transfer learning approaches are effective in overcoming data scarcity for training accurate predictive models in computational chemistry.
• These models are well-suited for integration into automated reaction mechanism generation and are expected to benefit other molecular property estimations.