The initial application of machine learning (ML) methodologies in the field of hydrological science surged toward the end of the 1990s (e.g., artificial neural networks, ASCE Task Committee, 2000), yet its usage remained limited primarily because researchers focused only on its predictive capabilities, mostly neglecting its potential for understanding underlying processes. ML research subsequently entered a hiatus period, not only in the field of hydrology, but also across all disciplines, likely due to data requirements and computational limitations.
Nearly a decade later, ML regained prominence, especially with the advent of the buzzword “deep learning,” achieving unprecedented success. In the first wave, ML was largely discredited as a “black box.” However, in the second wave, a newfound sense of awe emerged as entities across academic, non-academic, governmental, non-governmental, and startup sectors began amassing data to enhance learning capabilities.
Innovations like long short-term memory networks and generative techniques ignited such efforts in the field. In 2022, with the National Science Foundation’s (NSF) support, the HydroML Symposium was held at Pennsylvania State University, drawing more than 150 participants despite COVID-19-related challenges. Although ML was a significant theme across many conferences during this time, the focus was on enhancing hydrological response predictions, such as flood forecasting. By 2023, the emphasis had shifted beyond mere predictive capabilities, as seen in the second HydroML Symposium held at Lawrence Berkeley National Laboratory (LBNL). This successful event, aligned with the ExaSheds project, explored the integration of AI/ML with Earth System Science and drew more than 150 attendees. Subsequently, plans are underway for the next HydroML Symposium at Pacific Northwest National Laboratory (PNNL).
The primary insights from these two HydroML symposia as well as numerous other related events indicate that applying AI and ML in Earth System Science requires extensive research beyond simply predictive capabilities. Recent methodologies, such as operator learning (Xie et al., 2021; Li et al., 2020), differentiable modeling (Feng et al., 2023; Shen et al., 2023; Tsai et al., 2021), and interpretive AI (Mayer et al., 2021), have paved the way for significant advancements in this research field. Despite the challenges of the Earth system’s complexity and variable data availability, understanding causal inference and accurately representing processes are crucial steps. Moreover, a growing aspiration aims to blend mechanistic models with ML techniques.
To accomplish these goals, we have launched a new special collection titled “Advancing Interpretable AI/ML Methods for Deeper Insights and Mechanistic Understanding in Earth Sciences: Beyond Predictive Capabilities”. This is a joint special collection between Geophysical Research Letters, Water Resources Research, Earth’s Future, and JGR: Biogeosciences.
The scientific scope of the collection emphasizes enhancing ML techniques for better interpretability and predictive ability, developing AI/ML frameworks for Earth system diagnosis with sparse data, and employing ML for improved modeling and data-driven insights into Earth and its environmental systems. Submissions can include research letters, articles, reviews, methods, data papers, and commentaries as per the specific journal’s broader guidelines. Although the HydroML symposium initiated this special collection, we invite the participation of the entire hydrology and biogeosciences community worldwide. This special collection will serve as a platform propelling us into an era beyond mere predictive capabilities.
To submit your manuscript, use the submission site for Geophysical Research Letters, Water Resources Research, Earth’s Future, or JGR: Biogeosciences, and select the collection’s title from the drop down menu in the Special Collection field of the submission form. Manuscripts can be submitted to any of these journals depending on their fit with the journal’s scope and requirements.
—Dipankar Dwivedi (firstname.lastname@example.org, 0000-0003-1788-1900), Lawrence Berkeley National Laboratory, United States; Xingyuan Chen (0000-0003-1928-5555), Pacific Northwest National Laboratory, United States; Chaopeng Shen (0000-0002-0685-1901), Penn State University, United States; Harihar Rajaram (0000-0003-2040-358X), Johns Hopkins University, United States