Combining Group Contribution Method and Semisupervised Learning to Build Machine Learning Models for Predicting...

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Chemical Kinetics

Background:

• Machine learning (ML) models effectively predict reaction rate constants for organic compounds with hydroxyl radicals (HO•).
• Previous models faced limited applicability domains (AD) due to scarce experimental data (<1400 records).

Purpose of the Study:

• To address the limited AD of existing ML models for HO• reaction rate constants.
• To improve ML model performance and expand the AD by curating a larger dataset and employing data augmentation strategies.

Main Methods:

• Curated an expanded experimental dataset (Primary dataset) with 2358 kinetic records.
• Utilized group contribution method (GCM) and semisupervised learning (SSL) to augment the dataset.
• Incorporated 147,168 new data points into the final ML model.

Main Results:

• GCM enhanced model performance for chemicals outside the initial AD.
• SSL effectively expanded the model's AD.
• The final model achieved R² = 0.77, RMSE = 0.32, and MAE = 0.24 on the test set.
• The AD was expanded by 117%, enabling reliable predictions for over 560,000 chemicals.

Conclusions:

• The combined GCM and SSL approach effectively augments datasets, improving ML model performance and expanding AD.
• The developed ML model demonstrates reliable predictive capabilities for a vast chemical space.
• This research offers a robust method for enhancing ML models in chemical kinetics and provides a widely accessible online predictor.