Key points are not available for this paper at this time.
The applications of machine learning (ML) in hydrology have witnessed significant advancements in recent years. However, such applications have often occurred in relative isolation from the traditional mechanistic, process-based modeling (PBM) paradigms that have historically underpinned scientific discovery and policy support. This presentation contends that the cultural divide between the ML and PBM communities restricts the full potential of ML, even in its hybrid forms with PBM. A hydrologic modeling experiment is presented to illustrate the fundamental differences between these two perspectives and highlight critical yet overlooked challenges that ML may encounter in practice. These challenges stem from the inherent complexity of hydrologic systems, where behaviors can change in physically explainable ways not evident in historical records due to factors such as climate change and human interventions. The presentation explores a 'coevolutionary' model-building approach, advocating a shift from a borrowing culture to a co-creation culture. This shift aims to develop models that harness ML's strengths, such as scalability to big data and high-dimensional mapping, while remaining grounded in process-based knowledge and adhering to principles of model explainability, interpretability, and falsifiability. A novel modeling paradigm is proposed, one that is both ML-powered and process-equipped, facilitating knowledge discovery from vast, complex, and high-dimensional geospatial data. This paradigm enables the direct derivation and synthesis of new differential equations across various hydro-climatic and socio-economic settings, spanning scales from small headwater catchments to large multi-jurisdictional watersheds. The proposed modeling paradigm is evaluated through the simulation of rainfall-runoff mechanisms, with a specific focus on peak times, in diverse catchments across the Contiguous United States.
Razavi et al. (Mon,) studied this question.