Join Zoom Meeting
Meeting ID: 844 8406 0127 Passcode: 016243
https://illinois.zoom.us/j/84484060127?pwd=YSKwksbaq6ULaazkh2D8UX0CNa2kzH.1
Physically Informed Deep Learning for Multivariate River Solute Predictions
Advisor: Professor Praveen Kumar
Abstract:
River solute chemistry arises from complex, multiscale nonlinear interactions among
hydrological, geochemical, and landscape processes, often disrupted by human interventions
such as land use change and drainage network modifications. Traditional models capture either
rapid fluctuations or long-term shifts, but rarely both, limiting their utility in a dynamic
environmental context. This dissertation presents physics-informed machine learning models to
predict river solute concentrations across scales, using high-frequency data from three
contrasting catchments (Plynlimon in Wales, U.K., Orgeval in France, and Upper Sangamon in
Illinois, USA). We integrate machine learning models, Long Short-Term Memory (LSTM) and
Transformer architectures, with spatio-temporal data, including static catchment attributes
such as slope, porosity, aspect, and vegetation index (NDVI). First, we propose a novel
modification to the traditional gate model in LSTM, termed “flow-gate” to capture hysteresis.
Second, we develop an adaptive LSTM with additional gates, including “flux-gate” and “flowflux-
gate,” that activate based on flow and flux gradients to capture distinct regimes and
resulting solute dynamics. Third, we introduce the non-stationary Transformer with geospatial
features (NST-Geo), which uses static and NDVI embeddings to enhance long-term prediction
capabilities that can be attributed to spatial drivers. Together, these contributions improve the
interpretability and robustness of ML-based water quality predictions by integrating physical
insights.