Abstract:
River water quality is under heavy impact from anthropogenic activities. Organic anthropogenic pollutants that are present at low concentrations in river water, ranging from nanogram per liter to microgram per liter, are categorized as organic micropollutants. Traditional wastewater treatment plants (WWTPs) have limited capabilities to efficiently remove micropollutants before releasing treated water into the receiving rivers. WWTPs are one of the main point sources of micropollutants in rivers. Hence, the micropollutants from WWTPs directly impact the receiving waterbodies, particularly in small rivers where the surrounding areas are densely populated, and the WWTPs’ effluents make up a considerable amount of the river discharge during baseflow conditions.
The threat to river water quality from the presence of micropollutants in rivers is elevated during heavy rain events. Untreated wastewater enters rivers at larger quantities than during the baseflow conditions. Storms also provide micropollutants with various sources and dispersive entry routes into the receiving water. The level of impact from micropollutants is dictated by their fate in rivers. Therefore, when we evaluate, monitor, and improve river water quality, it becomes crucial to understand the source(s) and the in-stream reactive transport processes of micropollutants, which characterize their fate in rivers.
Micropollutants present in river water are composed of a wide range of substances, collectively referred to as micropollutant mixtures. It is impossible to identify every individual substance in the mixture. Previous studies used:
Chemical analysis to investigate the individually detected micropollutants in the mixture with regards to their concentrations.
In vitro bioassays to quantify the overall effects of the mixture with regards to their endpoints.
Numerical models to study the in-stream processes of single substances.
The goal of this thesis is to provide a quantitative understanding of the in-stream processes of the micropollutant mixture effects, which is still lacking in previous studies. To fill this knowledge gap, I hypothesized that the in-stream processes of the mixture effects are governed by the advection-dispersion-reaction equation (ADR). I tested the hypothesis by investigating the following three perspectives:
The in-stream processes of the mixture effects, for which I developed a convolution-based one-dimensional reactive transport model that is computationally cheap and is suitable to couple with Metropolis–Hastings Markov chain Monte Carlo algorithm for parameter estimates and computing ensemble model results. I parameterized the model to quantify the in-stream processes of individual micropollutants and their mixture in the Steinlach River near Tübingen, Southern Germany. The results show that our model parameterization can characterize the in-stream processes of the individual micropollutants and their mixture well. The low computational cost of the convolution enables modeling the fate of large numbers of substances, as well as many iterations of model runs during the Monte Carlo process.
The transferability of the ADR under different flow conditions for the mixture effects, for which I further developed a partial differential equation (PDE)-based one-dimensional transient reactive transport model and applied the model to a storm event, during which the mixture effects were sampled from the Ammer River, Tübingen, Southern Germany, and quantified in in vitro bioassays. I introduced stochastic elements into the model by using Gaussian Process Regression (GPR) to construct the model inputs. The conditional realizations from GPR enabled the model to efficiently generate ensemble results with deterministically calibrated parameter values while explicitly expressing the known physical processes. I showed GPR is a robust method to characterize the temporal in-stream dynamics of the mixture effects.
The potential to combine numerical approaches with machine learning methods to solve the ADR and estimate parameter values for the mixture effects, for which I applied the deep learning-based Bayesian optimization method, simulation-based inference (SBI), to obtain the parameter posterior distributions of a PDE-based one-dimensional reactive transport model. I also applied the physics-informed neural network (PINN) to obtain the solution of the same model. The two approaches were applied to model the in-stream processes of mixture effects sampled in the Ammer River during the baseflow conditions. I compared the results of modeled mixture using SBI coupled with the standard PDE model to those obtained with PINN. I showed the potential of deep learning methods aiding the process-based reactive transport modeling, demonstrating the advantages and disadvantages of both approaches, concluding that the preference of one over the other is a heavily objective-oriented choice.
To my knowledge, this is the first work that uses process-based models to quantitatively characterize the in-stream processes of the organic micropollutant mixture effects. My key results show the applicability of the mass conservation law to mixture effects quantified in in vitro bioassays, approving the validity of using mixture effects as a novel state variable for future water quality modeling, as well as highlighting the possibility and advantages of merging traditional process-based models with deep learning methods for future studies on mixture effects.