Where Machine Learning Fails in Predicting Emerging Contaminant Adsorption: A Decision-Oriented Framework for Model...

Area of Science:

• Environmental Science
• Computational Chemistry
• Data Science

Background:

• Emerging contaminants (ECs) pose environmental risks, making their adsorption behavior crucial for fate assessment and removal technologies.
• Machine learning (ML) models are increasingly used for predicting adsorption, but their reliability is hampered by workflow challenges.
• Existing research often overlooks systematic issues in ML model development, limiting practical application and transferability.

Purpose of the Study:

• To systematically review and address key challenges in constructing reliable ML models for EC adsorption prediction.
• To consolidate strategies for improving data quality, feature engineering, and model validation.
• To explore future research directions for advancing EC adsorption prediction.

Main Methods:

• Systematic literature review of ML applications in EC adsorption.
• Identification and categorization of challenges in ML model construction (data scarcity, bias, feature representation, applicability).
• Consolidation of improvement strategies including data quality assessment, automated feature extraction, cross-system modeling, and external validation.

Main Results:

• Identified key challenges: data scarcity/bias, low sample-to-feature ratio (SFR), inadequate feature representativeness, and poor model applicability/credibility.
• Proposed targeted solutions: rigorous data quality assessment, automated feature extraction, cross-system modeling, and external validation.
• Highlighted future directions: generative data augmentation, physics-informed ML (PIML), and automated literature data extraction using LLMs.

Conclusions:

• Addressing systematic challenges in the ML workflow is essential for robust and transferable EC adsorption prediction models.
• Implementing proposed strategies can enhance model reliability and credibility.
• Future research in areas like PIML and automated data extraction promises significant advancements in EC adsorption prediction and environmental applications.