AI Horizons
| What: | Twenty-two scientists gathered to explore how machine learning and artificial intelligence could transform Great Lakes observations, modeling, and forecasting. Participants discussed enhancing predictive models, improving observational network design, and identifying key challenges and opportunities. They aimed to establish a framework for a Great Lakes ML/AI community of practice and to develop a strategic vision for the next decade. |
| When: | 22–23 July 2024 |
| Where: | Ann Arbor, Michigan |
1. Introduction
The Michigan Institute for Data and AI in Society (MIDAS) and the Cooperative Institute for Great Lakes Research (CIGLR) hosted a summit in Ann Arbor, Michigan, from 22 to 23 July 2024. The event brought together 22 scientists from across the country to envision how machine learning and artificial intelligence might be used to address some of the most pressing challenges facing Great Lakes science, restoration, and management over the next decade.
2. Background
The Laurentian Great Lakes—comprising Lakes Superior, Michigan, Huron, Erie, and Ontario—are the largest group of freshwater lakes in the world by the total area, holding approximately 21% of Earth’s surface freshwater. In addition to their considerable size, the Great Lakes serve as a critical resource for over 40 million people in the United States and Canada. They provide drinking water, support commercial and recreational fishing, facilitate transportation and commerce, and offer opportunities for tourism, recreation, and connecting with nature. The Great Lakes also enrich regional biodiversity by supporting a wide range of interconnected ecosystems. Managed cooperatively by two neighboring countries, this dynamic system represents a case study for international water and ecosystem management and serves as a key testbed for understanding aquatic, terrestrial, and climatological systems globally.
Many different local, state, federal, and tribal government agencies, nonprofit organizations, academic institutions, private entities, and community stakeholders are involved in the complex scientific and management landscape of the Great Lakes. For example, the International Joint Commission (IJC) coordinates water management and water quality efforts between the United States and Canada, reflecting their mutual responsibilities for water use and safety. A variety of federal agencies, regional consortia, and national laboratories have developed regional, state-of-the-art monitoring networks that include buoys, vessel fleets, flux towers, and remote sensing capabilities. Notable organizations include the U.S. Environmental Protection Agency, Environment and Climate Change Canada, the National Oceanic and Atmospheric Administration’s (NOAA) Great Lakes Environmental Research Laboratory (GLERL), the Cooperative Institute for Great Lakes Research (CIGLR), and the Great Lakes Observing System (GLOS). Initiatives such as the Great Lakes Restoration Initiative, CoastWatch, the Synthesis, Observations, and Response System (SOAR), and Submerged Aquatic Vegetation (SAV) Mapping also play significant roles. The high concentration of observational networks in the region sets some parts of the Great Lakes apart as some of the best-monitored, large open water regions in the world. These efforts underscore the collaborative approach taken by multiple stakeholders to address pressing challenges facing the Great Lakes.
For example, despite considerable progress in recent years, accurate forecasting for the Great Lakes region remains a key challenge. Changes in lakewide water levels are set by a small residual of several large fluxes that can be difficult to forecast accurately (i.e., evaporation, precipitation, runoff). Even small errors in the estimates of these fluxes can lead to considerable errors in the water level forecast. As climate change continues to alter atmospheric thermodynamics and dynamics, projecting the net effect on the Great Lakes remains challenging. Evaporation tends to lower water levels and precipitation tends to raise them—both processes will be affected by climate change, and it is far from clear which one will dominate within a given forecast period. Accurate forecasting of ice conditions and harmful algal blooms (HABs) are also challenging.
Many Great Lakes datasets and studies are watershed specific and small scale, but understanding the entire Great Lakes—much like any large basin on Earth—requires a more holistic, integrated approach that considers the Great Lakes as a single system. Machine learning (ML) and artificial intelligence (AI), underpinned by the necessary data and infrastructure, present a unique opportunity for considerable advancements in forecasting, as well as other areas such as process-based understanding and observing network design.
Machine learning and artificial intelligence are transforming environmental science. For example, the field of oceanography has been revolutionized by ML/AI, both in terms of modeling capacity and scientific understanding (Sonnewald et al. 2021). Atmospheric science is similarly evolving, especially in the weather forecasting industry (Eyring et al. 2024). For example, the European Centre for Medium-Range Weather Forecasts (ECMWF) now runs a data-driven forecast system (Lang et al. 2024). ML/AI could also help improve relevant large-scale theories for hydrologic simulations and forecasting (Nearing et al. 2021). The Great Lakes research and management community should engage with these emerging toolkits in a coordinated and systematic manner to effectively achieve the greatest impact. Given the complexity of the science involved and the multifaceted nature of the Great Lakes community, we need a strategic approach.
3. CIGLR summit: AI horizons
The “AI Horizons” summit represents a strategic effort to envision the integration of ML/AI into Great Lakes science, management, and restoration. CIGLR Summits convene groups of 20–30 invited experts meeting for 2–3 days to summarize the state of knowledge and recommend future directions on Great Lakes problems that span decadal time scales. On 22 and 23 July 2024, CIGLR hosted a summit on the University of Michigan campus, specifically in the Samuel T. Dana Building, which is home to the School for Environment and Sustainability (SEAS). The summit brought together 22 researchers from across the country. The organizers identified two primary goals for the summit: creating a framework for a Great Lakes ML/AI community of practice and developing a decadal vision on how ML/AI could transform Great Lakes science. Over the 2-day summit, participants made progress on scaffolding and drafting up the vision document and engaged in foundational conversations on building a Great Lakes–centric community of practice.
The first day of the summit began with scene-setting talks. The organizers emphasized the importance of collaboration, stating that the success of the summit, like any collaborative endeavor, relies on the collective experience, energy, focus, and goodwill of the participants. To harness the collective intelligence of the participants, organizers implemented several tools and practices:
Shared, online notes documents that all summit participants could edit.
A group Zotero library, allowing for group management of relevant literature, citations, and bibliographies.
Dynamically defined working groups that evolved throughout the summit in response to participant feedback.
A “parking lot” for noting important topics that fell outside the formal agenda but warranted future discussion.
Web-based Q&A polls throughout the summit via the “Slido” tool.
Dedicated blocks of writing time, as part of the program, to allow for informal collaborative conversation, individual focus, and codevelopment of text to articulate ideas.
Jing Liu, the Executive Director of MIDAS, gave a brief overview of AI in science and engineering. She noted that data and AI are becoming essential tools in many domains, transforming both disciplinary and interdisciplinary research. This trend calls for new support mechanisms, such as new institutes. As one such institute, MIDAS is currently focused on enabling the adoption of AI methods in various research domains, promoting responsible and ethical AI, and exploring new ways of cross-sector collaboration with academia, industry, and government. Liu highlighted that the current academic research model has struggled to implement AI at the needed scale and pace to address pressing scientific problems, particularly in comparison to industry, where resources and talent tend to be concentrated.
In a second scene-setting talk, Dani Jones (CIGLR) echoed definitions of ML and AI as formulated by Liu. Specifically, “AI” refers to machines and/or algorithms that can perform tasks that normally require human intelligence, such as learning, reasoning, using and comprehending language, decision-making, and so on. Under that definition, “ML” may be considered a subset of AI focused on data-driven learning and prediction, which does have some overlap with traditional statistics.
Jones gave a brief overview of how machine learning has shaped the field of Earth system science. In oceanography, there have been considerable advances in subgrid scale parameterization and equation discovery (Bolton and Zanna 2019; Zanna and Bolton 2020), novel data analyses using unsupervised learning (Sonnewald et al. 2019, 2020), and hypothesis generation (Sonnewald et al. 2021). To set the stage to consider a framework that would lead into a Great Lakes–centric AI/ML community of practice, Jones shared examples of other centers of practice. In atmospheric science, organizations such as the Cooperative Institute for Research in the Atmosphere (CIRA) have pioneered the use of ML/AI in detection, inference, and estimation from satellite-based remote sensing products (Lagerquist and Ebert-Uphoff 2022). Large-scale initiatives and institutes like the NSF AI Institute for Research on Trustworthy AI in Weather, Climate, and Coastal Oceanography (McGovern et al. 2024) and the NASA JPL Science Understanding through Data Science (SUDS) initiative work to integrate ML/AI into Earth system sciences, emphasizing domain knowledge, trustworthiness, interpretability, and explainability.
Jones emphasized that Great Lakes science and management is just starting to use ML and is still in the early stages of conceptualizing how AI might be used. There have been ML-driven advances in predicting wave heights (Hu et al. 2021), ice cover (Abdelhady and Troy 2024), and water levels. There have also been advances in autonomous underwater sampling (Zhang et al. 2024) and model emulation and uncertainty quantification (Pringle et al. 2024). At present, there are efforts at GLERL and CIGLR to use data-driven methods to improve water level forecasting and improve understanding of the drivers of HABs. There are undoubtedly more efforts in the research community of which the authors are unaware, emphasizing the need for a central hub to coordinate these efforts and integrate ML/AI into Great Lakes science more effectively, supporting research, the exploration of new methods, training, and other relevant knowledge sets.
Dynamic working groups.
Before the summit, the organizers proposed the following working groups: 1) modeling and forecasting capacity, 2) observing network design, 3) data pipelines and data assimilation, and 4) hazards and risks. An initial survey revealed that interest in these working groups was unbalanced. The overwhelming majority of participants selected “modeling & forecasting capacity” as their first choice, while there was very little interest in “hazards and risks.” In response, the organizers split the modeling & forecasting capacity group into two: one focusing on predictive modeling and the other on operationalization and applications. The initially proposed “observing network design” and “data pipelines and data assimilation” groups were combined into a single group, and a new “generative AI” group was established. Participants then self-sorted into the working groups. During the morning working group discussions, participants focused on their respective topics. The groups highlighted issues, identified opportunities, and made recommendations to improve Great Lakes research and management through ML/AI:
Modeling and Forecast Capacity A: Predictive Models and Techniques. This group focused on pathways to develop and refine predictive models using machine learning, address technical challenges, enhance model accuracy, and explore innovative methodologies. Key discussion points included the following:
- ○Consideration of data from a statistical point of view: independence of samples, sufficiently large spatial datasets, and time series.
- ○Opportunities for linking highly time-resolved mechanistic models with less frequently sampled biological data to improve beach closures and understanding of harmful algal blooms. The group also emphasized leveraging physics-informed models while addressing the issue of “black box” models and various data characteristics depending on the application, such as dimensionality, sampling frequency, and predictive time scales.
- ○Challenges such as overfitting, enhancing model accuracy, and trade-offs between accuracy and generalizability.
Modeling and Forecast Capacity B: Operationalization and Applications. This group addressed the practical implementation of ML approaches in real-world scenarios, supporting decision-making processes and engaging stakeholders. Discussion points included the following:
- ○Forecasting at various time scales (short-term, subseasonal to seasonal, and years to decades).
- ○Using ML to better reconstruct past and current conditions (hindcast/nowcasting).
- ○Supporting large-scale infrastructure operation and policy with reinforcement learning and global approximators.
- ○Key challenges included stakeholder engagement and building trust while explaining and transparently communicating ML-based operations to the public and operators.
Generative AI: This group focused on exploring the potential of generative AI in Great Lakes science, restoration, and management. They discussed the following:
- ○The strengths and weaknesses of current technologies, such as integrating physics into cost functions.
- ○Potential for generative AI to fill in data gaps, considering that AI might not extrapolate well due to reliance on training datasets.
- ○Highlighting how physics-based optimization could help capture extremes and incorporate domain expertise.
Data Pipelines and Observing Network Design: This group looked at enhancing data integration and access, data ingestion pipelines, and mechanisms for AI to answer relevant questions for observation systems. They emphasized the following:
- ○The importance of accessible, robust, and analysis-ready data for ML/AI development.
- ○Utilizing AI for data curation, quality control, creating metadata, and supporting public decision-making through natural language processing.
- ○Addressing data sparsity, temporal issues, and ensuring accurate AI constraints by leveraging physics-informed models.
Following the morning working group breakout discussions, the participants reconvened to report on their discussions. It was decided that the working groups should be reorganized dynamically again by the participants before the afternoon session. The reorganized working groups aligned with the planned sections of the paper, specifically the following:
Philosophy, context, and problem formulation: Refining research questions, theoretical frameworks, and contextualizing challenges unique to the Great Lakes.
Benefits of AI in addressing Great Lakes challenges: Exploring specific applications and improvements brought by ML/AI.
Connecting with the wider landscape, best practices, and interoperability: Ensuring alignment with global initiatives and developing best practices.
The second day of the summit was dedicated to drafting the perspectives manuscript and developing the community of practice framework. Short breakout sessions allowed teams to refine sections of the manuscript, integrating insights from the previous day. Alternating between full-group discussions and focused writing sessions helped ensure alignment and coherence across different sections.
By late morning, outlines of key sections were ready for review. The generative AI group contributed insights into the strengths and weaknesses of current ML/AI technologies in meteorology, emphasizing the rapid growth in this discipline. They noted the potential for tailoring cost functions to capture physics better and using AI to extend data in data-sparse regions by leveraging physical laws. This approach motivates the collection of more data to enhance AI capabilities. Current technologies often struggle with capturing extremes due to issues with extrapolation, which is crucial for climate change projections. Physics-based optimizations can help incorporate domain expertise and potentially address these challenges.
A collaborative editing session followed, where participants provided feedback and integrated complementary ideas. The summit concluded with final remarks, reinforcing the importance of continued collaboration and outlining the next steps for the Great Lakes AI Laboratory and planned publications.
4. Summit outcomes
The summit aims to ultimately produce a perspectives paper on the effective integration of ML/AI into Great Lakes science and management. The work on the perspectives paper, which started during the summit, has continued post-summit, with participants engaging in writing groups and using collaborative tools to ensure coherence and comprehensive coverage of the identified themes and references.
a. Building up the Great Lakes AI community.
The second goal of the summit was to develop strategies, tools, and practices for fostering a collaborative AI research community in the Great Lakes region. As highlighted during the summit discussions, this effort should include the following:
A framework for a Great Lakes ML/AI community of practice: Establishing a structured approach to foster accessible, analysis-ready data [e.g., drawing inspiration from community initiatives like Pangeo (Odaka et al. 2020)] and reproducibility of both code and datasets. It was mentioned that the Great Lakes research community may not have the scale needed to create its own large-scale infrastructure; it may be more beneficial to join existing larger efforts, introducing a specific subfocus on the Great Lakes.
Community engagement: Promoting participation and collaboration across different stakeholders, leveraging diverse expertise and perspectives. This could take the form of sessions at national and international conferences (e.g., the American Geophysical Union’s Fall Meeting).
Balancing act: Addressing community-building efforts while mitigating concerns of communication burnout. This includes respecting the limited bandwidth of individuals and prioritizing sustainable development.
Support structures: Implementing sustainable support structures to maintain engagement and ensure continuous collaboration. This includes developing mechanisms for ongoing communication, resource sharing, and collaborative efforts.
b. Stakeholder engagement and trust building.
For successful integration, there must be a focus on engaging stakeholders and building trust in ML/AI systems. Transparent communication and explaining model outputs effectively are critical for gaining public and stakeholder trust.
5. Next steps
a. Perspectives manuscript.
Summit participants will draft and submit a perspectives article on the future of ML and AI in Great Lakes science and management. This manuscript, being developed by all summit attendees with additional expert input, will serve as a foundational document.
b. Launch of the Great Lakes AI laboratory.
The newly launched Great Lakes AI Laboratory will serve as a focal point for integrating ML and AI into Great Lakes research, management, and restoration, addressing the pressing need for advanced data pipelines, analytics, and predictive modeling in the region. Although CIGLR—consisting of a research institute and a Regional Consortium of academic, nongovernmental organization, and private sector partners—and NOAA GLERL will serve as the initial hosts for this initiative, the Great Lakes AI Laboratory is designed to be an open, collaborative community. We invite anyone with research interests or a stake in Great Lakes science, management, and restoration, as well as environmental machine learning experts looking to engage with the Great Lakes community, to join us. Specific platforms and events that the Great Lakes AI Laboratory will use may include a website, Slack channel, GitHub organization, mailing list, open workshops, and hackathons. By developing advanced predictive models, facilitating data integration, providing training and resources, and encouraging interdisciplinary collaboration, the Great Lakes AI Laboratory seeks to bridge the gap between data science and environmental science in the Great Lakes, fostering innovation and enhancing the region’s capacity to address its unique challenges. We invite interested parties to join us as part of the Great Lakes AI Laboratory GitHub organization (https://github.com/great-lakes-ai-lab).
Acknowledgments.
Funding for the summit was provided by the Cooperative Institute for Great Lakes Research (CIGLR), which is funded through the NOAA Cooperative Agreement with the University of Michigan (NA22OAR4320150). ADG (Gronewold) was partially supported through the NSF Global Centers program (NSF Award 2330317). This is NOAA GLERL Contribution Number 2067 and CIGLR Contribution Number 1254. The authors wish to thank all summit participants for their time, focus, and perspectives. We extend a special thanks to Margaret Throckmorton and Mary Ogdahl (CIGLR) and Beth Ubserseder (MIDAS) for their invaluable support in organizing and carrying out the meeting.
References
Abdelhady, H., and C. Troy, 2024: A deep learning approach for modeling and hindcasting Lake Michigan ice cover. arXiv, 2407.04937v1, https://doi.org/10.48550/arXiv.2407.04937.
Bolton, T., and L. Zanna, 2019: Applications of deep learning to ocean data inference and subgrid parameterization. J. Adv. Model. Earth Syst., 11, 376–399, https://doi.org/10.1029/2018MS001472.
Eyring, V., and Coauthors, 2024: Pushing the frontiers in climate modelling and analysis with machine learning. Nat. Climate Change, 14, 916–928, https://doi.org/10.1038/s41558-024-02095-y.
Hu, H., A. J. van der Westhuysen, P. Chu, and A. Fujisaki-Manome, 2021: Predicting Lake Erie wave heights and periods using XGBoost and LSTM. Ocean Modell., 164, 101832, https://doi.org/10.1016/j.ocemod.2021.101832.
Lagerquist, R., and I. Ebert-Uphoff, 2022: Can we integrate spatial verification methods into neural network loss functions for atmospheric science? Artif. Intell. Earth Syst., 1, e220021, https://doi.org/10.1175/AIES-D-22-0021.1.
Lang, S., and Coauthors, 2024: AIFS – ECMWF’s data-driven forecasting system. arXiv, 2406.01465v2, https://doi.org/10.48550/arXiv.2406.01465.
McGovern, A., and Coauthors, 2024: AI2ES: The NSF AI Institute for research on trustworthy AI for weather, climate, and coastal oceanography. AI Mag., 45, 105–110, https://doi.org/10.1002/aaai.12160.
Nearing, G. S., F. Kratzert, A. K. Sampson, C. S. Pelissier, D. Klotz, J. M. Frame, C. Prieto, and H. V. Gupta, 2021: What role does hydrological science play in the age of machine learning? Water Resour. Res., 57, e2020WR028091, https://doi.org/10.1029/2020WR028091.
Odaka, T. E., A. Banihirwe, G. Eynard-Bontemps, A. Ponte, G. Maze, K. Paul, J. Baker, and R. Abernathey, 2020: The Pangeo Ecosystem: Interactive computing tools for the geosciences: Benchmarking on HPC. Tools and Techniques for High Performance Computing, G. Juckeland and S. Chandrasekaran, Eds., Springer International Publishing, 190–204.
Pringle, W. J., and Coauthors, 2024: Coupled lake-atmosphere-land physics uncertainties in a Great Lakes regional climate mode. ESS Open Archive, 171052472.24971474/v1, https://doi.org/10.22541/essoar.171052472.24971474/v1.
Sonnewald, M., C. Wunsch, and P. Heimbach, 2019: Unsupervised learning reveals geography of global ocean dynamical regions. Earth Space Sci., 6, 784–794, https://doi.org/10.1029/2018EA000519.
Sonnewald, M., S. Dutkiewicz, C. Hill, and G. Forget, 2020: Elucidating ecological complexity: Unsupervised learning determines global marine eco-provinces. Sci. Adv., 6, eaay4740, https://doi.org/10.1126/sciadv.aay4740.
Sonnewald, M., R. Lguensat, D. C. Jones, P. D. Dueben, J. Brajard, and V. Balaji, 2021: Bridging observations, theory and numerical simulation of the ocean using machine learning. Environ. Res. Lett., 16, 073008, https://doi.org/10.1088/1748-9326/ac0eb0.
Zanna, L., and T. Bolton, 2020: Data‐driven equation discovery of ocean mesoscale closures. Geophys. Res. Lett., 47, e2020GL088376, https://doi.org/10.1029/2020GL088376.
Zhang, Y., and Coauthors, 2024: Using a long‐range autonomous underwater vehicle to find and sample harmful algal blooms in Lake Erie. Limnol. Oceanogr. Methods, 22, 473–483, https://doi.org/10.1002/lom3.10621.
