Rapid Prediction of Buoyancy-Driven Exchange Flows at the Great Salt Lake: ML Models and a 1D Shallow Water Approach

The Great Salt Lake in Utah, USA, is a hypersaline terminal lake divided in to northern and southern arms by the Union Pacific Railroad causeway since the 1950's. This separation has caused a difference in density and water surface elevation between lake arms. These differences result in a buoyancy-driven exchange flow occurring through an engineered breach in the causeway. Traditionally, modeling the flow through the breach has been done by numerically solving the 1D steady shallow water equations, and using computational fluid dynamics (CFD). The CFD models yield high accuracy results, but require substantial computing resources. This research proposes the use of data measured by United States Geological Survey (USGS) to create data-driven models to predict the exchange flow through the breach. The use of data-driven models, often referred to as machine learning, allows for faster flow prediction and requires lower computational cost compared to CFD simulations. This study uses, Linear Regression, Random Forests, Support Vector Regression, and Deep Neural Networks to create data-driven models from available USGS data. These models are compared to physics-based prediction models and monthly measurements taken by USGS. The results of this study show that data-driven models can accurately predict the buoyancy-driven exchange flow at a time consistent with USGS' sampling. These models could serve as a method for real-time prediction of the flow through the breach in the Great Salt Lake, facilitating better management of the flow between the arms of the lake and informing changes to the lake conditions over time.