Machine Learning Applications for Chemical Fingerprinting and Environmental Source Tracking Using Non-target Chemical...

Area of Science:

• Environmental Chemistry
• Forensic Science
• Machine Learning Applications

Background:

• Chemical fingerprinting is crucial for identifying pollution sources in environmental samples.
• Existing chemical fingerprinting workflows are often qualitative, limiting precise source attribution.
• A quantitative approach is needed for accurate and sensitive detection of chemical sources.

Purpose of the Study:

• To develop and validate a quantitative machine learning workflow for chemical source identification.
• To assess the workflow's ability to select diagnostic chemical features for predicting source presence.
• To evaluate the workflow's performance in detecting sources at trace levels and high dilutions.

Main Methods:

• Collected 51 grab samples from five distinct chemical sources: agricultural runoff, headwaters, livestock manure, (sub)urban runoff, and municipal wastewater.
• Employed support vector classification to identify top discriminating chemical features (10, 25, 50, 100) for each source.
• Validated the workflow using cross-validation (92-100% balanced accuracy) and simulated in silico mixtures.

Main Results:

• The machine learning workflow achieved high cross-validation accuracy (92-100%) in identifying chemical sources.
• Screening environmental samples revealed low presence probabilities for most sources, except for wastewater at specific downstream locations.
• The workflow successfully distinguished the presence/absence of some sources at 10,000-fold dilutions in simulated mixtures.

Conclusions:

• The developed quantitative workflow effectively selects diagnostic chemical features for source prediction.
• This method enables the quantitative prediction of various environmental sources at trace levels.
• The workflow shows promise for sensitive and accurate environmental monitoring and forensic analysis.