Chaopeng Shen
Associate ProfessorDepartment of Civil and Environmental Engineering
Pennsylvania State University
Tuesday, November 7 at 1:00-2:00 PM PT / 4:00-5:00 PM ET
ABSTRACT
Process-based modelling offers interpretability and physical consistency in many domains of geosciences but struggles to leverage large datasets efficiently. Machine-learning methods, especially deep networks, have strong predictive skills yet are unable to answer specific scientific questions. A recently developed genre of physics-informed machine learning, called “differentiable” models, connect neural networks (NNs) with process-based equations (priors) to benefit from the best of both NNs and process-based modeling paradigms. We propose that differentiable models are especially suitable as global hydrologic models (GHMs) because they can harvest information from big earth observations to produce state-of-the-art predictions, enable physical interpretation naturally, extrapolate well (due to physical constraints) in space and time, and leverage progress in modern AI computing architecture and foundation models. Differentiable models can also learn from the experiences of and synergize with existing GHMs, providing support to the GHM community. Differentiable GHMs can answer pressing societal questions on water resources availability, climate change impact assessment, and disaster risk mitigation, among others. Development of differentiable GHM shows it can approach or even exceed the performance level of deep learning models in some of the aspects, e.g., representation of extremes or under data-sparse scenarios, but we can learn a great about processes because it allows us to distinguish better priors. We demonstrate the power of differentiable modeling using computational examples in rainfall-runoff modeling, river routing, as well applications in water-related domains such as ecosystem modeling and water quality modeling. Furthermore, we show how differentiable modeling can enable us to ask fundamental questions in geosciences and get robust answers from big global data.
Presentation Slides
Video Link (Coming soon!)