Machine Learning-Based Identification of Petroleum Distillates and Gasoline Traces Using Measured and Synthetic GC...

Area of Science:

• Forensic Science
• Analytical Chemistry
• Computer Science

Background:

• Arson investigations rely on forensic experts analyzing gas chromatography (GC) spectra to detect ignitable liquids.
• Current methods are time-consuming, especially when ignitable liquids are absent.
• Machine Learning (ML) offers a potential solution to automate and improve the accuracy of spectral analysis.

Purpose of the Study:

• To develop and validate an ML-based workflow for classifying ignitable liquids using GC chromatograms (spectra).
• To create a novel spectra synthesis algorithm to augment limited experimental data.
• To evaluate the impact of dataset size and ML algorithms on classification performance.

Main Methods:

• Collected and annotated 181 real spectra from fire scenes and reference databases.
• Developed ML models using k-Nearest Neighbors (kNN), representative spectrum, and Random Forest (RF) algorithms.
• Created a spectra synthesis algorithm to generate a large dataset of synthetic spectra.
• Trained and tested kNN, RF, representative spectrum, and Deep Learning (DL) models on both real and synthetic datasets.

Main Results:

• ML models achieved high F1-scores (0.74-0.96) on real spectra when trained on augmented datasets.
• Performance was more dependent on the size of the training dataset than the specific ML algorithm used.
• Models trained on larger, synthesized datasets showed improved predictive power on independent real spectra.

Conclusions:

• The developed ML workflow and spectra synthesis algorithm effectively classify ignitable liquids from GC spectra.
• Increased dataset size significantly enhances ML model performance in forensic spectral analysis.
• The methodology is adaptable to other forensic domains utilizing spectral data.