Application of machine learning and deep learning methods for hydrated electron rate constant prediction.

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Chemical Kinetics

Background:

• Accurate determination of the second-order rate constant with hydrated electrons (k_eaq^-) is essential for understanding organic compound degradation in advanced reduction processes (ARPs).
• Existing methods for predicting k_eaq^- often face challenges with accuracy and over-fitting, particularly for diverse organic compounds (OCs).

Purpose of the Study:

• To develop and compare machine learning (ML) and deep learning (DL) models for predicting k_eaq^- of OCs.
• To evaluate the performance of a convolutional neural network (CNN) with transfer learning and data augmentation (CNN-TL&DA) against traditional ML algorithms.
• To interpret the models to identify key molecular descriptors and environmental factors influencing k_eaq^-.

Main Methods:

• Collected 867 k_eaq^- values from peer-reviewed literature across various pH conditions.
• Applied XGBoost (ML) with Mordred descriptors (MD) and Morgan fingerprints (MF) for prediction.
• Utilized a CNN model incorporating transfer learning and data augmentation (CNN-TL&DA) with molecular images (MI).
• Employed SHAP for interpreting ML models and Grad-CAM for interpreting CNN models.

Main Results:

• The CNN-TL&DA model achieved superior prediction performance (R²_test = 0.896, RMSE_test = 0.362, MAE_test = 0.261), significantly outperforming XGBoost with MD (R²_test = 0.692) and MF (R²_test = 0.512).
• Model interpretation revealed that molecular size, branching, electron distribution, polarizability, bond types, functional groups (e.g., -CN, -NO₂, -X), and pH are critical factors influencing k_eaq^-.
• CNN models effectively identified key functional groups, particularly electron-withdrawing groups, contributing to higher k_eaq^- values.

Conclusions:

• The CNN-TL&DA approach demonstrates a significant improvement in predicting k_eaq^- for OCs, effectively overcoming over-fitting issues.
• CNN models exhibit lower prediction errors compared to traditional ML algorithms, highlighting their potential for predicting other rate constants in environmental chemistry.
• This study provides a robust computational framework for assessing the reactivity of organic compounds in ARPs, aiding in the design of more efficient remediation strategies.