A framework for developing machine learning-based chemical fingerprinting models using large gas chromatograph-mass...

Area of Science:

• Environmental Forensics
• Analytical Chemistry
• Machine Learning

Background:

• Chemical fingerprinting is crucial for oil spill investigations, traditionally relying on gas chromatograph-mass spectrometer (GC-MS) analysis of biomarkers.
• Identifying oil sources involves complex analysis of biomarker distributions and diagnostic ratios, which can be challenging with weathered samples.

Purpose of the Study:

• To develop a machine learning (ML) framework for direct chemical fingerprinting of oil spills from GC-MS data.
• To overcome limitations of conventional methods, such as the need for biomarker identification and the scarcity of weathered oil samples.

Main Methods:

• A novel ML framework was developed, utilizing pattern recognition directly on GC-MS data.
• Methodological innovations include training classifiers on fresh oil data, generating synthetic data, and using aggregate ensemble models.
• Standardized preprocessing ensured data compatibility across different GC-MS instruments.

Main Results:

• The ML framework successfully performed chemical fingerprinting without identifying individual biomarkers or calculating diagnostic ratios.
• Aggregate ensemble models trained exclusively on fresh, unweathered crude oils demonstrated high accuracy in classifying spill residues and identifying source oils.
• The approach proved robust across various sample types, overcoming analytical limitations.

Conclusions:

• This study presents a novel, practical ML-based approach for oil spill chemical fingerprinting.
• The framework enhances environmental forensics investigations by providing accurate and efficient source identification.
• The method offers a significant advancement over traditional biomarker analysis techniques.