A primary objective of environmental forensics is associating contaminant impacts in environmental media to potential sources. Experienced practitioners can recognize patterns in contaminant distributions (i.e., fingerprinting), to make such associations. However, these analyses are time consuming and subject to bias. This work significantly reduces these disadvantages through the application of machine-learning algorithms.  A Random Forest model was trained on polycyclic aromatic hydrocarbon (PAH) sediment fingerprints compiled from 13 sites throughout the United States. Human practitioners examined each fingerprint and identified one of three dominant PAH patterns – heavy pyrogenic, light pyrogenic, and petrogenic – and the model was trained on the labeled fingerprint data. Overall, the trained Random Forest classifier concurred with the human practitioners in 92% of the test set sediment samples, with the majority of practitioner/model disagreement occurring in the ambiguous space between heavy and light pyrogenic categories. The trained model correctly categorized 100% of pure phase reference samples. No difference in model performance was observed between sites. This modeling approach can be used at future sites and the method is generalizable to other compound classes. This machine learning application will improve the efficiency and defensibility of environmental forensic evaluations. One promising approach is identifying samples for which the model is confident in its classification, allowing a human practitioner to focus further efforts on the remaining, more ambiguous samples (typically 10-20% of a given sample set).

Multi-site machine learning-enhanced source classification of PAH impacted sediments: Environmental forensics applications

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.