49 - Enhancing Modeling of Drinking Water Quality with Physics Informed Neural Networks (PINNs): Learning from incomplete Reaction Model and Incomplete Data

Chemical kinetics models, mathematically represented by a system of ordinary or partial differential equations (ODEs or PDEs), are useful tools to simulate chemical reactions relevant to many environmental settings, including drinking water quality. However, these models can be inaccurate due to: (1) the laboratory conditions under which the model was developed do not match the real-world conditions to which the model is being applied and (2) limitations in experimental analytical methods might limit the description of the true underlying chemical mechanisms involved. Further, traditional methods of solving ODEs or PDEs lack the ability to assimilate data to improve the accuracy of model predictions. To address these limitations, this work proposes a Physics Informed Neural Network (PINN) to improve the prediction of chemical formation/decay of multiple species over time by accounting for a known (but inaccurate) chemical model and experimental data. Using reactions describing water disinfectant and disinfection byproduct formation that are critical for public health and regulatory compliance, we show that the PINN model is able to accurately predict the concentrations of chemical species for different pH values, and for chemical species for which data was not available during model training. While the original motivation of this work is to accurately predict chemical concentrations relevant to drinking water disinfection, the method and framework developed in this paper can be applied to many chemical systems, yielding a wide array of potential applications in a variety of domains.