Introduction

Machine learning and artificial intelligence (ML/AI) techniques have advanced considerably across many domains and applications in recent years, and the field of Earth system science is increasingly utilizing ML/AI tools. Earth system analysis and prediction has been a major user and driver of advanced computing dating back to the mid-20th century, when early weather and climate models were first built. Simulations of Earth system processes—for example, weather, climate, biosphere, hydrologic, and geophysical processes—have routinely pushed capabilities of computing systems to their limits; today, models and observing systems are being linked to examine Earth system-wide problems and interactions among different parts of the system. Interest is increasing in near- to long-term forecasts of the Earth system at higher spatiotemporal resolution to support decision making.

Future progress in advancing scientific understanding and in providing reliable and high-quality Earth system science forecasts and data products to a wide range of users will require access to adequate computing resources and use of cutting-edge analytical capabilities. Several trends in Earth system simulation—for example, the growing complexity of science models aimed at improving accuracy, the desire for higher resolution simulations, the use of ensembles for uncertainty quantification, and the incorporation of growing volumes of data from new observing platforms—all point to the importance of more powerful and efficient tools and techniques, and expanded computing and analytical capabilities.

ML/AI offer the potential to advance capabilities for Earth system observing, modeling, analysis, prediction, understanding, and decision making in various ways, including optimizing automated simulation workflows, providing new ways to simulate and predict the behavior of physical systems, and improving parameterizations of physical and chemical processes. While the Earth system science community has made good use of many statistical and modeling methods, ML/AI techniques could help transform difficult computational problems into those that can better exploit ongoing advances in information frameworks and computer architectures optimized for ML/AI computation.

In addition to these opportunities and emerging approaches for using ML/AI to advance Earth system science, there are a number of challenges and risks including technical and data challenges, such as the interoperability of data and the integration of ML/AI tools with existing infrastructure; necessary workforce development; and the “last mile” problem of making forecasts useful for a wide array of users and decision makers. Developing and using the algorithms ethically includes acknowledging the implicit biases and historical lack of diversity in Earth system science communities: biases in human experts and ML/AI outputs may have implications for the use and acceptance of ML/AI tools and applications. The evolving nature of the public-private-academic interface has also created a complicated landscape for the application of ML/AI in Earth system science. While the government was formerly the only entity that could support and maintain large observing and modeling systems, the paradigm is shifting as data, models, and computing resources are becoming more accessible to many. Cloud-based providers are beginning to cultivate scientists and federal agencies as customers especially around data storage and analysis, and increasingly for some types of computationally intensive workloads. Private companies and nongovernmental organizations are launching a next generation of observing systems, generating large amounts of proprietary data to be used for analysis and prediction.

In February 2022, the Board on Atmospheric Sciences and Climate together with the Board on Mathematical Sciences and Analytics, the Computer Science and Telecommunications Board, the Board on Earth Sciences and Resources, and the Ocean Studies Board of the National Academies of Sciences, Engineering, and Medicine convened a workshop on the opportunities and challenges of using ML/AI to advance Earth system science, including their ethical development and use. The workshop convened Earth system science experts, ML/AI researchers, social and behavioral scientists, ethicists, and decision makers across sectors to explore how these approaches can contribute to improving understanding, analysis, modeling, prediction, and decision making. Key terms used throughout the proceedings are defined in Box 1. See Appendix A for the statement of task to the workshop planning committee; see Appendix B for biographical sketches of the planning committee members; and see Appendix C for the workshop agenda.

The structure of the proceedings largely follows the workshop agenda (see Appendix C). The 3 days of the workshop were organized around 3 broad themes:

1. Emerging approaches for using, interpreting, and integrating ML/AI for Earth system science;
2. Challenges and risks of using ML/AI for Earth system science; and
3. Future opportunities to accelerate progress.

This proceedings summarizes workshop presentations and discussions in the plenary sessions. This proceedings has been prepared by the workshop rapporteur as a factual summary of what occurred at the workshop. The planning committee’s role was limited to planning and convening the workshop. Funding for this workshop was provided by the National Aeronautics and Space Administration (NASA), the National Oceanic and Atmospheric Administration (NOAA), and the National Science Foundation (NSF) through the core sponsorship provided to the National Academies’ Board on Atmospheric Sciences and Climate.

OVERVIEW OF STATE-OF-THE-ART USE OF ML/AI FOR EARTH SYSTEM SCIENCE

Peter Dueben, European Centre for Medium-Range Weather Forecasts (ECMWF), provided a broad overview of how ML/AI approaches are currently being used for Earth system science, including conceptual and technical challenges for using ML/AI and opportunities to address those challenges. The Earth system is a large, complex system that is chaotic and not linear with many interacting components, making it difficult to simulate with models. At the same time, between both observational and model data, a rich set of information is available about the Earth system. ML is an important tool for Earth system science, Dueben explained, because ML has the power to extract information from existing data and make sense of it in areas where conventional models struggle (Box 2).

Dueben identified four high-level categories of how ML/AI have been or currently are used in Earth system science: (1) improve understanding of the system (e.g., fusing of information from different sources, causal discovery, uncertainty quantification); (2) speed up simulations and green computing (e.g., emulation of model components, data compression); (3) improve models (e.g., learn components from observations, correct biases); and (4) link communities (e.g., health, energy, transportation). Using the ECMWF workflow as an example, Dueben emphasized that rather than a single ML/AI workflow, there are many ML/AI applications throughout the numerical weather prediction workflow, from observations to data assimilation and numerical weather prediction forward modeling and post-processing.

Dueben provided a vision for making the most of ML/AI in the mid-term future. He suggested getting organized as a community as a first step, and he showed the 10-year roadmap that was developed at ECMWF as an example (Figure 1; Dueben et al., 2021). In the United States, NOAA has developed an AI strategic plan (NOAA, 2021), and NSF is in the process of doing

the same.³ As a next step, Dueben suggested designing software to make it easy for domain scientists to get started. For example, CliMetLab is an effort to share boilerplate code so that domain scientists can focus on the ML code (Raoult and Pinault, 2022). In order to advance science, Dueben identified the need to build ML tools applicable to the Earth science domain (e.g., Adewoyin et al., 2021), learn how to use ML at scale (e.g., the MAELSTROM project⁴), and combine operational models with ML (e.g., Laloyaux et al., 2022).

Additionally, in order to facilitate innovation and open science, Dueben suggested making ML developments comparable using benchmark datasets specific to the Earth sciences. Benchmark datasets are useful because they allow quantitative evaluation of ML approaches, access to relevant data, and separation of concerns between domain scientists and ML and high-performance computing experts. Ideal benchmark datasets would include a problem statement; data that are available online; python code or Jupyter⁵ notebooks; a reference ML solution; quantitative evaluation metrics; visualization, diagnostics, and robustness tests; and computational benchmarks.

___________________

³ See https://www.nsf.gov/cise/ai.jsp

⁴ See https://www.maelstrom-eurohpc.eu

⁵ See https://jupyter.org

This page intentionally left blank.