SDS Colloquium, Speaker Andrew Bennett

Title: Building harmony between process-based and machine-learning models for hydrologic
prediction

Abstract: Machine-learning (ML) based models are both popular and state-of-the-art in terms
of predictive performance for hydrologic modeling. However, these models are often not very
flexible once trained and are generally not usable beyond their targeted variables and
spatiotemporal scales. On the other hand, process-based (PB) hydrologic models are flexible in
their spatiotemporal resolutions and often simulate multiple processes together, making them
extremely useful for scientific understanding and operational predictions across a wide range of
target areas such as energy, agriculture, and flood forecasting. In this talk I will highlight my
work in reconciling these approaches. I will discuss topics such as hybrid modeling, model
emulation, differentiable hydrology, and how the ongoing explosion in data availability is
changing the state of hydrologic predictions.

Relevant Papers:

• Deep Learned Process Parameterizations Provide Better Representations of Turbulent Heat Fluxes in Hydrologic Models
• Spatio-Temporal Machine Learning for Regional to Continental Scale Terrestrial Hydrology