Using three-dimensional fluorescence and artificial neural networks to match diesel-contaminated ground water with possible sources

Petroleum products released into the environment pose a health threat because of the toxic and carcinogenic hydrocarbons they contain. This study utilizes 3-dimensional fluorescence spectroscopy and artificial neural networks to help identify the source of diesel fuel that appeared in a freshly excavated sewer trench in Havre, Montana. Persistent polycyclic aromatic hydrocarbons (PAH's) naturally occurring in diesel provide a fluorescent fingerprint that is characteristic of a given source.; Samples of petroleum contaminated ground water were collected from the sewer trench, from wells at five potential source areas within the Havre rail yards and from several monitoring wells upgradient from the sewer trench. Three source areas were of major interest because of their proximity to the trench: an open lagoon holding wastewater from the rail yard operations, a fueling station near the locomotive maintenance shops, and an adjacent recovery zone from which subsurface diesel is actively being recovered. Monitoring wells from each area showed significant diesel fuel contamination. Our technique attempts to determine which of the three suspect areas, or combination thereof, is responsible for the interceptor trench contamination.; In the course of these analyses, a systematic description of concentration effects in fluorophore mixtures (the red shift cascade) is offered. The concentration-imposed complications are demonstrated for a synthetic mixture mimicking diesel and several ternary PAH mixtures.; Samples from the Havre field site were extracted into benzene and filtered through 0.45mm PTFE membranes. The extracted samples were then characterized by 3-dimensional fluorescence spectroscopy which collects an array of 8421 measurements over excitation and emission wavelength ranges of 200 to 600 nm. Seven sequential dilutions of each extract were also measured to capture the entire range of spectral information available from the red shift cascade effect.; Matching of contaminant spectra to source spectra was performed with artificial neural networks. Digitized fluorescence data from the original extract plus the seven dilutions were concatenated into a single fact file. Fact files from the 3 known areas within the rail yard site were combined with some creosote samples and random files to train the network. Several fact files were withheld from the training run and used to test the network's ability to recognize samples from each source area. Seven generations of network design are chronicled. Finally, contaminant samples of unknown source origin were submitted to the best trained and tested network for assignment of source area probabilities. Results correlated well with observations from hydrological engineers studying the same site.