Physics Informed Neural Network for 1,4 Dioxane Plume

The Gelman plume is a 1,4-dioxane mass spreading through Ann Arbor's groundwater. Dioxane is a regulated carcinogen in Michigan and concerned residents + Michigan's Environment, Great Lakes, and Energy (EGLE) department have tracked it for decades. The plume is on track to reach Huron River, which is where 85% of the City of Ann Arbor's drinking water is sourced from. There is currently no dioxane treatment method at the drinking water treatment plant. As of March 2026, Ann Arbor is the United States' newest Superfund site due to the plume; the classification means the federal U.S Environmental Protection Agency (EPA) will organize increased mitigation measures and resources for the site. Standard discretized ground water transport models of the site (eg. MODFLOW) need flow and dispersion parameters supplied up front, which are expensive to measure and often poorly constrained.

I built a physics-informed neural network that embeds the advection-dispersion equation directly in its loss function, so the model fits decades of monitoring-well data while staying consistent with the physics of how contaminants actually move. Rather than taking transport parameters as inputs, it infers them during training.

The model reaches R² = 0.974 against the EGLE's monitoring data and recovers a plume velocity of 169.6 ft/yr, a dispersivity of 12 ft, and a first-order decay rate of 0.018/yr, none of which it was given. These parameters match established published modeling work done for the plume.

I presented my work to both the U.S. EPA and to the Ann Arbor Drinking Water Treatment plant for further collaboration and my work's use as a decision support tool for both organizations. It is also deployed as a public Streamlit tool so others can explore the predictions (Gelman PINN Forecaster). A v4 is in progress, adding a time-varying source boundary condition and a 10-model ensemble for uncertainty bounds.