Formatted Title:

Using a Machine Learning Model to Predict Chemical Sources at Former Manufactured Gas Plant Sites

Background/Objectives:

Polycyclic aromatic hydrocarbons (PAHs), along with other chemicals, are used as benchmarks in assessing complex former manufactured gas plant (MGP) sites. For example, the ratio of fluoranthene to pyrene (FL/PY) can be used to distinguish between pyrogenic and petrogenic sources: FL/PY ratios less than 1 may indicate the presence of petroleum, and FL/PY ratios greater than 1 may indicate pyrolytic origins associated with combustion or high-temperature pyrolysis. Recently, there has been interest in applying machine learning approaches for source attribution at contaminated sites. Machine learning algorithms, such as random forest, can be used to develop models that predict categorical chemical characteristics at target sites based on "training" data from similar, well-studied sites. This study uses random forest models to predict whether the FL/PY ratio of a given environmental sample from a target MGP site is above or below 1 based on publicly available information from other MGP site investigations. Potentially, this model can be expanded, for example, to predict whether chemical concentrations at a given MGP site (or contaminated sites) exceed regulatory screening criteria.

Approach/Activities:

Random forest is a supervised machine learning algorithm based on the construction of multiple decision trees during model training. As a classification tool, the model output is 1 or 0, where 1 represents a FL/PY ratio above 1 and 0 represents an FL/PY ratio below 1. The 1 or 0 output is determined based on the highest agreement rate among the "n" decision trees. We built random forest classification models using publicly available data from various former MGP site investigations in the northeastern United States (US). We constructed the model with FL/PY ratios as the target variable, and concentration data of various other chemicals, spatial data, and historical use information as predictor variables. The initial training dataset was created from a wide variety of MGP sites, but was subsequently filtered based on the model’s variable importance score, a model attribute that quantifies the influence of each predictor variable on the model output. Variable importance scores were generated using a training/test split method, in which the dataset was randomly split – 70% was used as the model training set and 30% was used as the model test set. Once refined, the training set became an amalgamation of MGP sites sharing similar topography, size, and access to surface water that could be used as a surrogate to predict the FL/PY ratio at other test sites.

Results/Lessons Learned:

While variable importance scores varied across different models, the highest ranked importance factors were concentrations of PAHs and volatile petroleum hydrocarbons (VPHs), and location-based variables, such as ground surface and sample elevation. Refined models had FL/PY prediction accuracy rates between 84% and 93%. These models gained approximately two-thirds of their predictive power from the top one-fifth of predictor variables, sorted by importance score. This study introduces machine learning methods to accurately predict source attribution characteristics at former MGP sites, and it tests whether random forest models can be applied to sites with limited data (e.g., poorly studied sites). It is important to note that the predictive power of the random forest model is limited by the data upon which the tool is trained, and that datasets must be refined continuously based on available (albeit limited) information to improve the model’s ability to predict target characteristics. Using random forest models to predict FL/PY ratios shows that machine learning algorithms, like random forest, can be used to predict various categorical attributes of target MGP sites based on groups of surrogate sites.

Samuel Schrage

Affiliation

Refine Results

(Group 2, Poster Board #28) Using a Machine Learning Model to Predict Chemical Sources at Former Manufactured Gas Plant Sites