Modeling dynamic geophysical phenomena is at the core of Earth and environmental studies. The geoscientific community relying mainly on physical representations may want to consider much deeper adoption of artificial intelligence (AI) instruments in the context of AI's global success and emergence of big Earth data. A new perspective of using hybrid physics‐AI approaches is a grand vision, but actualizing such approaches remains an open question in geoscience. This study develops a general approach to improving AI geoscientific awareness, wherein physical approaches such as temporal dynamic geoscientific models are included as special recurrent neural layers in a deep learning architecture. The illustrative case of runoff modeling across the conterminous United States demonstrates that the physics‐aware DL model has enhanced prediction accuracy, robust transferability, and good intelligence for inferring unobserved processes. This study represents a firm step toward realizing the vision of tackling Earth system challenges by physics‐AI integration.

Long short‐term memory (LSTM) networks represent one of the most prevalent deep learning (DL) architectures in current hydrological modeling, but they remain black boxes from which process understanding can hardly be obtained. This study aims to demonstrate the potential of interpretive DL in gaining scientific insights using flood prediction across the contiguous United States (CONUS) as a case study. Two interpretation methods were adopted to decipher the machine‐captured patterns and inner workings of LSTM networks. The DL interpretation by the expected gradients method revealed three distinct input‐output relationships learned by LSTM‐based runoff models in 160 individual catchments. These relationships correspond to three flood‐inducing mechanisms—snowmelt, recent rainfall, and historical rainfall—that account for 10.1%, 60.9%, and 29.0% of the 20,908 flow peaks identified from the data set, respectively. Single flooding mechanisms dominate 70.7% of the investigated catchments (11.9% snowmelt‐dominated, 34.4% recent rainfall‐dominated, and 24.4% historical rainfall‐dominated mechanisms), and the remaining 29.3% have mixed mechanisms. The spatial variability in the dominant mechanisms reflects the catchments' geographic and climatic conditions. Moreover, the additive decomposition method unveils how the LSTM network behaves differently in retaining and discarding information when emulating different types of floods. Information from inputs within previous time steps can be partially stored in the memory of LSTM networks to predict snowmelt‐induced and historical rainfall‐induced floods, while for recent rainfall‐induced floods, only recent information is retained. Overall, this study provides a new perspective for understanding hydrological processes and extremes and demonstrates the prospect of artificial intelligence‐assisted scientific discovery in the future.

Interpretable Machine Learning (IML) has rapidly advanced in recent years, offering new opportunities to improve our understanding of the complex Earth system. IML goes beyond conventional machine learning by not only making predictions but also seeking to elucidate the reasoning behind those predictions. The combination of predictive power and enhanced transparency makes IML a promising approach for uncovering relationships in data that may be overlooked by traditional analysis. Despite its potential, the broader implications for the field have yet to be fully appreciated. Meanwhile, the rapid proliferation of IML, still in its early stages, has been accompanied by instances of careless application. In response to these challenges, this paper focuses on how IML can effectively and appropriately aid geoscientists in advancing process understanding—areas that are often underexplored in more technical discussions of IML. Specifically, we identify pragmatic application scenarios for IML in typical geoscientific studies, such as quantifying relationships in specific contexts, generating hypotheses about potential mechanisms, and evaluating process‐based models. Moreover, we present a general and practical workflow for using IML to address specific research questions. In particular, we identify several critical and common pitfalls in the use of IML that can lead to misleading conclusions, and propose corresponding good practices. Our goal is to facilitate a broader, yet more careful and thoughtful integration of IML into Earth science research, positioning it as a valuable data science tool capable of enhancing our current understanding of the Earth system.

Climate change may systematically impact hydrometeorological processes and their interactions, resulting in changes in flooding mechanisms. Identifying such changes is important for flood forecasting and projection. Currently, there is a lack of observational evidence regarding trends in flooding mechanisms in Europe, which requires reliable methods to disentangle emerging patterns from the complex interactions between flood drivers. Recently, numerous studies have demonstrated the skill of machine learning (ML) for predictions in hydrology, e.g., for predicting river discharge based on its relationship with meteorological drivers. The relationship, if explained properly, may provide us with new insights into hydrological processes. Here, by using a novel explainable ML framework, combined with cluster analysis, we identify three primary patterns that drive 53 968 annual maximum discharge events in around a thousand European catchments. The patterns can be associated with three catchment-wide river flooding mechanisms: recent precipitation, antecedent precipitation (i.e., excessive soil moisture), and snowmelt. The results indicate that over half of the studied catchments are controlled by a combination of the above mechanisms, especially recent precipitation in combination with excessive soil moisture, which is the dominant mechanism in one-third of the catchments. Over the past 70 years, significant changes in the dominant flooding mechanisms have been detected within a number of European catchments. Generally, the number of snowmelt-induced floods has decreased significantly, whereas floods driven by recent precipitation have increased. The detected changes in flooding mechanisms are consistent with the expected climate change responses, and we highlight the risks associated with the resulting impact on flooding seasonality and magnitude. Overall, the study offers a new perspective on understanding changes in weather and climate extreme events by using explainable ML and demonstrates the prospect of future scientific discoveries supported by artificial intelligence.

Estimating river flood risks under climate change is challenging, largely due to the interacting and combined influences of various flood-generating drivers. However, a more detailed quantitative analysis of such compounding effects and the implications of their interplay remains underexplored on a large scale. Here, we use explainable machine learning to disentangle compounding effects between drivers and quantify their importance for different flood magnitudes across thousands of catchments worldwide. Our findings demonstrate the ubiquity of compounding effects in many floods. Their importance often increases with flood magnitude, but the strength of this increase varies on the basis of catchment conditions. Traditional flood analysis might underestimate extreme flood hazards in catchments where the contribution of compounding effects strongly varies with flood magnitude. Overall, our study highlights the need to carefully incorporate compounding effects in flood risk assessment to improve estimates of extreme floods.

While deep learning (DL) models exhibit superior simulation accuracy over traditional distributed hydrological models (DHMs), their main limitations lie in opacity and the absence of underlying physical mechanisms. The pursuit of synergies between DL and DHMs is an engaging research domain, yet a definitive roadmap remains elusive. In this study, a novel framework that seamlessly integrates a process‐based hydrological model encoded as a neural network (NN), an additional NN for mapping spatially distributed and physically meaningful parameters from watershed attributes, and NN‐based replacement models representing inadequately understood processes is developed. Multi‐source observations are used as training data, and the framework is fully differentiable, enabling fast parameter tuning by backpropagation. A hybrid DL model of the Amazon Basin (∼6 × 10 ⁶ km ² ) was established based on the framework, and HydroPy, a global‐scale DHM, was encoded as its physical backbone. Trained simultaneously with streamflow observations and Gravity Recovery and Climate Experiment satellite data, the hybrid model yielded median Nash‐Sutcliffe efficiencies of 0.83 and 0.77 for dynamic and distributed simulations of streamflow and total water storage, respectively, 41% and 35% higher than those of the original HydroPy model. Replacing the original Penman‒Monteith formulation in HydroPy with a replacement NN produces more plausible potential evapotranspiration (PET) estimates, and unravels the spatial pattern of PET in this giant basin. The NN used for parameterization was interpreted to identify the factors controlling the spatial variability in key parameters. Overall, this study lays out a feasible technical roadmap for distributed hydrological modeling in the big data era.