A Hybrid process-guided machine learning approach for estimating groundwater dynamics

To better understand groundwater dynamics in response to inner and external forces is essential for sustainable development and management. Traditionally, a wide array of process-based conceptual and numerical models are main tools for flow simulation. The process-based models can precisely capture the complexity and details of the hydrological cycle and physical processes. There is a tendency of constructing larger, finer and more complex distributed models due to management needs. However, these models are challenged by data scarcity, spatiotemporal scale, computational efficiency and simulation uncertainty. As data-driven models are involving rapidly, they can predict groundwater levels without deep knowledge of underlying hydrogeological properties, and recognize patterns hidden in collected data. But machine learning models lack of consideration for hydrogeological heterogeneity and spatiotemporal relationships between hydrogeological components. In the study, we used the hybrid process-guided machine learning approach that leverages strengths of hypothesis-driven physical model and data science. A well-constructed hybrid model speeded up complex models without sacrificing accuracy or detail. Most importantly, hybrid models are computationally efficient and could be transferrable across basins.