Improved Machine Learning Models by Data Processing for Predicting Life-Cycle Environmental Impacts of Chemicals

Area of Science:

• Environmental Science
• Computational Chemistry
• Chemical Engineering

Background:

• Machine learning (ML) offers efficient prediction of chemical life-cycle environmental impacts.
• Current ML models face challenges with low prediction accuracy and poor interpretability.
• Improving these models is crucial for reliable environmental assessments.

Purpose of the Study:

• To enhance the accuracy and interpretability of ML models for predicting chemical life-cycle environmental impacts.
• To develop improved data processing methods for ML in environmental science.
• To validate the effectiveness of novel feature and data selection techniques.

Main Methods:

• Implemented a mutual information-permutation importance (MI-PI) feature selection method to filter molecular descriptors.
• Applied a weighted Euclidean distance method for relevant data mining and feature contribution quantification.
• Developed artificial neural network (ANN) models using processed data.

Main Results:

• Achieved high R-squared values for ANN models across various environmental impact categories (e.g., 0.86 for terrestrial acidification, 0.84 for metal depletion).
• The MI-PI and weighted Euclidean distance methods improved model accuracy and interpretability.
• Shapley additive explanation confirmed the contribution of molecular descriptors to predictions.

Conclusions:

• The combined approach of MI-PI feature selection and weighted Euclidean distance data selection significantly enhances ML model accuracy and interpretability.
• This methodology shows promise for more reliable prediction of chemical life-cycle environmental impacts.
• The study provides a robust framework for developing interpretable ML models in environmental risk assessment.