Machine Learning-Assisted QSAR Models on Contaminant Reactivity Toward Four Oxidants: Combining Small Data Sets and...

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Machine Learning Applications

Background:

• Predicting organic contaminant reactivity with oxidants is crucial for environmental remediation.
• Small sample sizes in experimental data pose challenges for developing accurate predictive models.
• Existing machine learning (ML) models often struggle with limited data for oxidant reactivity prediction.

Purpose of the Study:

• To develop and evaluate methods for improving predictive models of organic contaminant reactivity.
• To investigate the effectiveness of combining small data sets and knowledge transfer for ML models.
• To compare proposed approaches with multitask learning and image-based transfer learning.

Main Methods:

• Merged small data sets into a unified dataset for training a single ML model.
• Developed knowledge transfer models between pairs of small data sets.
• Evaluated model performance using root-mean-square error (RMSE) on test sets for four oxidants: sulfate radical (SO4•−), hypochlorous acid (HClO), ozone (O3), and chlorine dioxide (ClO2).

Main Results:

• The unified ML model demonstrated improved predictive performance across all tested oxidants compared to individual models.
• Knowledge transfer models showed varied improvements, depending on data set consistency and individual model performance.
• The proposed methods consistently outperformed multitask learning and image-based transfer learning in enhancing predictive accuracy.

Conclusions:

• Combining small, similar data sets effectively improves ML model performance for predicting oxidant reactivity.
• Knowledge transfer between data sets can enhance predictive capabilities, contingent on shared knowledge and baseline model performance.
• The study validates novel approaches for building robust predictive models in environmental science, particularly with limited data.