Speakers
Travis Swanson
Applied computational earth scientist. My domain expertise is in computational sedimentary geology, but my passion and capacity to lead and support interdisciplinary projects is evidenced by my prior collaborative work on topics that include geomorphology, sedimentology, petroleum geology, ecohydrology, physical and chemical hydrology, and planetary science.
Presentation Description
A Physics-Based and Machine-Learning Hybrid Model Supporting the RESTORE-funded Lowermost Mississippi River Management ProgramThe Lowermost Mississippi River Management Program aims to evaluate benefits and costs of management schemes for the Lowermost Mississippi River (LMR), including unintended adverse impacts under a range of projected futures. Doing so requires a modeling approach that accurately predicts water flow, sediment transport, and channel bed evolution and sediment storage over the entire LMR for long time scales (50-100 yrs), and that can be efficiently applied to obtain results for a large range of scenarios, alternatives, and combinations. The physics of water flow is well understood and predictable over these space and time scales. The same cannot be said for sediment transport, where physics-based models typically exhibit major uncertainties and limitations or accumulate error when used over long time scales or complex spatial domains. Machine Learning (ML) approaches offer the potential for models that are fast, accurate over complex spatial domains, and are well suited to deal with unusual influences on sediment transport such as hysteresis in the upstream sediment supply or anthropogenic influences like intensive dredging on ad hoc schedules. We will demonstrate a hybrid modeling approach wherein a validated USACE HEC-RAS sediment transport model is used to train suites of ML models of sediment transport throughout the LMR. The ML models are validated against modeled and observational data and are then used in place of the HEC-RAS sediment transport model, while still being driven by HEC-RAS flow outputs. This approach, known as surrogate modeling, is proven in numerical weather prediction , where the physics is also complex and fast run times are a priority, but has not yet been widely adopted by the geomorphology community. We will demonstrate the effectiveness and utility of this approach at key locations in the LMR.