Previous Chapter: 4 Living with Wildland Fire: AI to Inform Adaptation
Suggested Citation: "5 AI for Water Resource Management." National Academies of Sciences, Engineering, and Medicine. 2026. Accelerating Climate Progress with AI: From Science to Action: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/29400.

5

AI for Water Resource Management

The use of artificial intelligence (AI) to advance climate action related to hydrology, water quality, water resources, and water resource management was explored in this panel, alongside challenges to using AI-enhanced tools to try to clarify uncertainty in these areas. Adrienne Wootten, University of Oklahoma, moderated a panel discussion examining the integration of AI with hydrologic and climate modeling and potential AI use for management decision making. The panelists were Debaditya Chakraborty, University of Texas–San Antonio, and Kathy Boomer, Foundation for Food and Agricultural Research.

PANELIST REMARKS

Bridging Physics, Data, and Decision Making

Chakraborty spoke about opportunities to increase AI trustworthiness and effectively apply AI tools to water resource management. His team works on explainable and counterfactual AI, which clarifies why a system makes a particular prediction and can uncover patterns in the data to reveal insights that can be useful for decision making. For example, they are working to use AI to quantify the impacts of climate interventions and adaptation and mitigation strategies to move beyond correlation to causation. They are also identifying patterns to find out, for example, which regions within a basin would benefit most from conservation.

He pointed out that for structured, region-specific data, where accuracy and interpretability matter, simpler and more transparent models may outperform black-box AI approaches. Additionally, relying on opaque AI

Suggested Citation: "5 AI for Water Resource Management." National Academies of Sciences, Engineering, and Medicine. 2026. Accelerating Climate Progress with AI: From Science to Action: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/29400.

outputs can lead to poor or risky decisions. He highlighted the risk of negative outcomes when tools are deployed before proper generalization testing and adequate understanding of why the model makes specific predictions.

In terms of model selection, Chakraborty said that his team’s research has shown that for skewed datasets, choosing slightly simpler models rather than more complex models like deep learning convolutional neural networks or transformers leads to better predictions and more meaningful physical outcomes. He also suggested that applying several layers of tests, rather than just one, on top of the AI-based results can lead to more trustworthiness and ultimately better adoption. His team’s comparison of simpler models to deep learning for short-term predictions of water levels or spring flows showed that simple models performed better when predicting 1–12 weeks out.

Chakraborty emphasized that it is important to not only look at the point predictions but also fully examine the usefulness of the output when selecting the best tool for a given situation. For example, does the model predict the extreme values, such as the drought periods or peak recharge instances? Complexity does not equal accuracy, he cautioned, noting that his team has found oversmoothing to be a critical problem with deep learning. He said that it may be necessary to take extreme care when using skewed data with very complicated black-box models. It is also instructive to use counterfactual AI when investigating climate impacts, to predict what would happen if a specific intervention of interest were not implemented.

Closing the Gap Between Technology and Implementation

While a lot of useful field technology has been developed, Boomer said that limited implementation is one of the main challenges to ensuring water and food security for future generations. As an example, she noted that advanced technologies, such as precision irrigation and conservation drainage, are implemented in less than 5 percent of crop lands in the United States and around the world.

This gap between technology availability and implementation comes down to trust, she said. Often, stakeholders, land managers, and farmers either do not trust or do not see the relevance of a technology or source of information for making decisions on their fields or farm operations. One of the most important reasons for this gap is that the people who are generating the information often do not adequately understand the intended decision context. Boomer noted an increasing recognition that trust-building comes from integrating local knowledge with academic knowledge rather than disseminating information to end users. Recent research in social science, economics, and decision making shows the power of more collaborative approaches, and Boomer said that such approaches naturally align

Suggested Citation: "5 AI for Water Resource Management." National Academies of Sciences, Engineering, and Medicine. 2026. Accelerating Climate Progress with AI: From Science to Action: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/29400.

with Indigenous knowledge systems, which have long emphasized shared responsibility and collective decision making.

She noted a well-established consultation framework that provides a process for facilitating collaboration. Effectively using this framework requires trust, space for reciprocity, valuing everybody in the conversation, and finding ways to incorporate local knowledge by creating space for feedback and having the willingness to adjust times and modeling approaches to think about trade-offs and best solutions. She suggested that AI could be used to fast-track trust-building. For example, AI could be valuable for cataloging the different ideas brought up during the problem-framing stage, including capturing ideas, concerns, and constraints around mitigation strategies and then synthesizing or integrating those diverse perspectives in an equitable way to foster community alignment and consent.

The assessment phase in the consultation framework provides an opportunity for quantitative comparison of management alternatives using science-based modeling tools. For improving water model performance, Boomer noted great potential to use AI to help with pre- and postprocessing of model data, facilitating model implementation (e.g., calibration, validation, and sensitivity analyses), supporting multimodeling approaches, and postprocessing model output into usable information. This will help scientists develop models that are more relevant to decision making and characterize uncertainty or risk (e.g., decomposing structural, parameter, and observational uncertainty), she said. Further, assessment with AI will reveal where, when, and what data are needed to strengthen decisions.

Boomer also highlighted that consultation is value based, so AI cannot replace humans, only support the process. She further acknowledged that enabling consultation requires a trained workforce. In the problem-framing and reflection phases of the consultation framework, elicitation and facilitation skills are essential to supporting authentic dialogues that can draw out ideas while maintaining connections to the broader aspects of the framework. In the assessment phase, she highlighted the importance of modeling fundamentals, including an understanding of the diverse range of models and their capabilities (e.g., conceptual, physical, empirical/statistical, processes-based, and hybrid models) and training in modeling best practices to support consultation assessments. For action and evaluation of the consultation framework, she highlighted the need for a workforce that can translate model results to practitioners implementing actions on the ground and data collectors evaluating system conditions and action outcomes. Finally, Boomer underscored the importance of recognizing the limitations of all models, including AI, across all phases of this consultation framework. These barriers include skills to ensure transparency, identify spurious results, and manage context shifts.

Suggested Citation: "5 AI for Water Resource Management." National Academies of Sciences, Engineering, and Medicine. 2026. Accelerating Climate Progress with AI: From Science to Action: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/29400.

She pointed out that Indigenous knowledge systems offer a critical conceptual model centered on the principle that water is life, recognizing how ground- and surface-water flows over daily, seasonal, and longer time scales shapes landforms, ecosystems, agricultural capacity, and fire risk. She contrasted this with many AI models, which are often driven mainly by precipitation data and may fail to capture the full complexity of water systems. She cautioned that ignoring these alternative conceptual models increases the risk of poor decisions by overlooking hidden costs in land and water management, potentially undermining capacity for evaluating threats to water and food security. She called for water-centered thinking, knowledge co-development, greater emphasis on conceptual modeling, and multiple and nested models—rather than a single “best” model, supported by AI—to better account for uncertainty, spatial and temporal change, and drift over time. She emphasized that smart modeling, rather than more data, is essential for improving decision making and managing computational demand.

PANEL DISCUSSION

Defining and Building Trust

Panelists explored the issue of trust during the discussion. Boomer said that trust can be thought of as a need to surface and value different ideas about how the socioecological system is behaving and can be managed. Finding uncertainties and confronting them together is an exciting opportunity to foster collaboration on the ground while also advancing research. It is also critical to test or field-validate the information generated using AI and other modeling tools in the decision-making context to evaluate if it is useful and makes sense, from both the end user and science perspectives (e.g., how the system behaves and might respond to management practices).

Chakraborty pointed out that issues around trust in AI have been discussed for years. From a data science point of view, trustworthiness comes from knowing what generalization tests have been performed to ensure that the AI is not just memorizing patterns but actually learning and providing an answer based on a rational interpretation of the dataset it was trained on. However, newer questions of trust tend to be more focused on how holistic the dataset is.

Reflecting Communities’ Experience

Monica Morrison, National Science Foundation National Center for Atmospheric Research, noted that much of the data currently used fail to capture phenomena that matter to the communities that might eventually use

Suggested Citation: "5 AI for Water Resource Management." National Academies of Sciences, Engineering, and Medicine. 2026. Accelerating Climate Progress with AI: From Science to Action: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/29400.

these tools. She suggested this gap extends beyond data scarcity and might occur because climate and Earth system data are unusual in that they do not explicitly describe phenomena; instead, meaning can be extracted through interpretation. She raised the question of how, particularly for hydroclimate and hydrological phenomena, data can be more rigorously interrogated to ensure they reflect the community experience—and what implications this has for the data processes used to train and evaluate models.

Noting that applied research is research that informs a decision, Boomer said that it is key to not only understand the needs of decision makers but also engage them in developing the scenario assessments that might be applied to a model under development. Chakraborty agreed with Morrison that data scarcity is challenging and said that generative AI and specific models, such as diffusion models, autoencoders, and latent space perturbations, look promising for overcoming data scarcity problems. However, it is still challenging to determine the trustworthiness of the datasets generated using such models.

Boomer noted that data scarcity will always be an issue because conditions change over time, and data will never be sufficient to capture what happened in the past in a way that can be used to predict new conditions. This underscores the need for process-based models and points to a potential weakness of AI models. She suggested that hybrid and multimodeling approaches provide the best strategy for assessing conditions and evaluating mitigation options.

Improving Drought Monitoring

In response to a question about how to use AI to improve drought monitoring, assessment, and forecasting, Chakraborty highlighted how skewed classification makes it difficult to use historical datasets to predict droughts. Increasing the data used for underrepresented information in a dataset (minority oversampling) and relying less on the more common data (majority undersampling) can be used to overcome this problem, along with diffusion models which are showing some potential for generating synthetic data points, especially for extreme droughts and peak recharge, he noted. He also emphasized that scale matters: focusing on regional rather than global scales would likely yield more precise and actionable results. Wootten noted that a similar problem arises in the context of predicting flooding and extreme rain events.

Boomer highlighted the general need for more robust approaches to understanding uncertainty in data and decomposing uncertainty in a way that reveals where more information can be useful. Chakraborty noted that local explanations are especially valuable when trying to figure out why a model is predicting a drought, for example, and whether it is correct.

Suggested Citation: "5 AI for Water Resource Management." National Academies of Sciences, Engineering, and Medicine. 2026. Accelerating Climate Progress with AI: From Science to Action: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/29400.

The plausibility of those inflection-point explanations can then be examined by a domain expert to enhance the model’s robustness and validity.

Enhancing Communication with AI

An online participant asked about pitfalls and benefits to incorporating AI into communicating across disciplinary boundaries and perspectives. Boomer expressed enthusiasm regarding its capacity to help various groups understand each other’s vocabularies and identify alignments and differences in views of the system. This could surface uncertainties and foster collaboration, she said. Chakraborty agreed that AI is useful in communication and underscored the continued need for human expertise. He said that more AI and domain experts are needed, as well as educators who can help with properly communicating the usefulness of AI and the best ways to use it.

Suggested Citation: "5 AI for Water Resource Management." National Academies of Sciences, Engineering, and Medicine. 2026. Accelerating Climate Progress with AI: From Science to Action: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/29400.
Page 37
Suggested Citation: "5 AI for Water Resource Management." National Academies of Sciences, Engineering, and Medicine. 2026. Accelerating Climate Progress with AI: From Science to Action: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/29400.
Page 38
Suggested Citation: "5 AI for Water Resource Management." National Academies of Sciences, Engineering, and Medicine. 2026. Accelerating Climate Progress with AI: From Science to Action: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/29400.
Page 39
Suggested Citation: "5 AI for Water Resource Management." National Academies of Sciences, Engineering, and Medicine. 2026. Accelerating Climate Progress with AI: From Science to Action: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/29400.
Page 40
Suggested Citation: "5 AI for Water Resource Management." National Academies of Sciences, Engineering, and Medicine. 2026. Accelerating Climate Progress with AI: From Science to Action: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/29400.
Page 41
Suggested Citation: "5 AI for Water Resource Management." National Academies of Sciences, Engineering, and Medicine. 2026. Accelerating Climate Progress with AI: From Science to Action: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/29400.
Page 42
Next Chapter: 6 Use of AI in Agriculture and Land Management
Subscribe to Emails from the National Academies
Stay up to date on activities, publications, and events by subscribing to email updates.