Developing a Machine-Learned Bayesian Network Model to Screen for VOC Vapor Intrusion

Abstract Description: The movement of underground volatile organic compounds into indoor air, known as vapor intrusion, can increase the risk of adverse health outcomes particularly harmful to unborn children and pregnant mothers, such as preterm birth, low birth weight, cancer, and more. These dangers are especially prevalent in less affluent and minority communities, who already have poorer access to medical care. For these groups especially, vapor intrusion may seem unavoidable – and it is widely under-addressed due to high costs of monitoring. Currently, the U.S. Environmental Protection Agency (EPA) recommends the Johnson–Ettinger Model for screening: a deterministic, one-dimensional, and steady-state diffusion model using static parameters to predict vapor intrusion risk. However, because many Johnson–Ettinger Model input parameters are expensive and impractical to collect, default parameters are often used, making it less sensitive to site-specific factors. This project aims to develop statistical and machine learning screening models to remove data barriers to identifying households in most need of mitigation. A machine-learned Bayesian network model is being trained on a nationwide EPA vapor intrusion database augmented with the Johnson–Ettinger Model. A Bayesian network is a graphical probabilistic model especially promising for capturing relationships among variables and modeling with incomplete datasets. Additionally, a probabilistic version of the Johnson–Ettinger model is being trained using the same database. This version accounts for the probability distributions of each input parameter in the Johnson–Ettinger model and uses Monte Carlo simulations to estimate the mean and uncertainty of the model predictions. The Bayesian network’s accuracy in predicting homes at risk from PCE and TCE vapor intrusion will be compared to deterministic and probabilistic versions of the Johnson–Ettinger model. Preliminary results indicate that the BN model is capable of achieving high accuracy in predicting indoor vapor intrusion exposure for both PCE and TCE. These results will help local public health and environmental agencies identify homes at risk and in need of support to decrease vapor intrusion risks.