Identifying Future Opportunities to Accelerate Progress
With the increasing interest in and adoption of ML/AI approaches to advance Earth system science, there are opportunities for disruptions to lead to rapid scientific, societal, or institutional progress in the Earth system science field. Ultimately, the interactions between scientific advances and the needs of decision makers—both individual users and policymakers—can enrich the development of operational tools, and questions remain about how decision makers could be engaged in the development of ML/AI tools for Earth system science applications. In addition, funding and structures that can facilitate the exchange of ideas underlie any vision for progress. Participants discussed opportunities to consider new models that could enable convergent science, knowledge transfer, and partnerships to support science at the intersection of ML/AI and Earth system science.
USING ML/AI FOR DATA-DRIVEN DECISION MAKING
Speakers in this session discussed the role of ML/AI at the interface of predictive physical models and real-time decision making. Discussions in this session also considered how to handle uncertainties in applications where mistakes can be costly, and the remaining science, engineering, societal, and ethical challenges of using ML/AI for data-driven decision making. Each speaker considered a particular case study through which they examined these critical questions.
ML Applications for Improved Earthquake Hazard Decision Making
Elizabeth Cochran, U.S. Geological Survey (USGS), outlined the many different earthquake-related products that USGS provides to aid in decision making by emergency managers, government officials, and the public including long-term forecasts before an earthquake, earthquake early warning, and products looking at the impacts and potential aftershocks after an earthquake happens (Figure 11). Currently, state-of-the-art real-time applications of ML are used for earthquake detection. Sensors distributed across the United States and globally provide a time series of ground motion—if that motion is an earthquake, then scientists can tell how big the earthquake was, as well as where and what time the earthquake occurred. Historically, this work was done by hand; starting several decades ago, computer algorithms were created to automate this process, but these algorithms were not very reliable. Over the past decade, ML approaches have been developed using deep neutral networks to identify seismic phases (Ross et al., 2018). Cochran explained that these algorithms, built off high-quality hand-curated datasets, have yielded great results. ML tools can be powerfully applied to increasingly large datasets from sensors, telecommunications fiber optic cables, and cell phones (Allen, 2012; Brooks et al., 2021; Cochran, 2018; Lindsey et al., 2020). Furthermore, ML tools used for earthquake detection also allow scientists to better understand the physics of
earthquake processes—for example, detecting and understanding very small events (e.g., Ross et al., 2020), and revealing fault zones where critical infrastructure should not be built.
Earthquake early warning—sending alerts to areas likely to experience damaging shaking—is a challenging problem that requires quickly detecting an event and its evolution. ML classifiers for signal versus noise detection can remove signals that are not of interest (e.g., Li et al., 2018; Meier et al., 2019) and can improve location estimates using real-time identification (e.g., Mousavi et al., 2020; Ross et al., 2018). Graph neural networks can incorporate information across a seismic network rather than just from a single station, which could allow for estimating source information and predicting shaking amplitudes (van den Ende and Ampuero, 2020).
Additional ML applications for earthquake information include ShakeMap, which aims to quickly identify the distribution of shaking (Kubo et al., 2020); PAGER, which looks at building responses and other impacts from ground motions (Xie et al., 2020); and improved earthquake forecasts to predict the time of the next earthquake (Rouet-Leduc et al., 2017). Remaining challenges include overcoming the lack of high-quality datasets and potentially biased data for complex earthquake hazard applications (e.g., seismic fragility, forecasting), and the consideration of benefits versus costs in the integration of these ML methods into an operational system.
Challenges in Employing ML/AI for Nowcasting
Daniel Rothenberg, Waymo, focused his remarks on examples related to nowcasts—short-term weather forecasts primarily focused on the next few hours—to examine the role that AI plays in data-driven decision making. Nowcasts are generally focused on a specific set of weather hazards or impacts with the potential to impact people, a place, or an asset, and are strongly constrained by current or recent weather conditions (e.g., tornado warnings, precipitation nowcasts). Classical nowcasting approaches (Figure 12) are highly data driven and rely on several approaches: radar extrapolation techniques that use computer vision to analyze patterns of motions and currents and historical radar imageries to extrapolate them into the
future to predict near-term storm patterns; identification of specific storm cells or hazards that are apparent in radar imagery and project where they are likely to move over the short term; and full complexity simulations of the atmosphere, including numerical weather prediction, to create short-term forecasts.
Precipitation nowcasting has been a target for early applications of AI in the weather domain because of the apparent simplicity of the problem from a technical perspective—predicting future images given a sequence of recent radar imagery (e.g., Klocek et al., 2021; Ravuri et al., 2021; Sønderby et al., 2020). Nowcasting involves pattern recognition and extrapolation from recent data, which makes it a good problem for well-established AI technologies that typically excel at image processing, segmentation, and classification. AI can augment and complement the capabilities of more traditional data-driven approaches. Rothenberg noted that nontraditional stakeholders in the weather and climate enterprise have driven many of these advances.
Rothenberg asserted that weather data and weather forecasts have no intrinsic value; rather, the processes that lead to decisions to protect life and property are where value creation happens. The utility of an AI tool at the interface of traditional physical modeling in support of a decision-making process is tied to whether or not the AI tool can improve the outcome of that process in a meaningful way. Rothenberg described the limitations of using AI for precipitation nowcasts. For example, specialized AI tools have been developed to represent nonlinear weather changes (e.g., convective initiation), but AI-based precipitation nowcasts generally do not have this capability. Additionally, most AI applications in nowcasting still heavily rely on numerical weather prediction systems to augment data for training or at time-of-inference, which introduces a question of whether AI-based systems are worthy of further investment or if classical techniques should continue to receive investments.
Open challenges remain in applying AI to precipitation nowcasting. AI requires a tremendous amount of high-quality data, though there is a lack of common datasets for benchmarking (Veillette et al., 2020), and, as other participants noted, not all data are equitably distributed geographically, which may introduce bias to the decision-making process. Rothenberg emphasized that if AI is the future of precipitation nowcasting, then it would be critical to move beyond the data-rich regions of the United States and Europe and develop solutions for poorly
observed parts of the world. As a first step, the community could determine whether existing data (e.g., existing satellite constellations) could be used if they were more accessible, or if new technologies such as space-borne radars would need to be launched.
Learning Dynamical Models, Uncertainty Quantification, and Intelligent Sampling and Autonomy
Pierre Lermusiaux, Massachusetts Institute of Technology, outlined several dynamical model learning approaches, from a more traditional Earth science case in which prior information is known from long-term datasets and knowledge of a model prior, to approaches for which there is no prior knowledge and ML tools estimate model equations and parameters (Figure 13). In between these two approaches is the growing field of scientific ML, where data-driven, explainable models can incorporate prior domain knowledge.
Lermusiaux considered examples of learning bottom gravity current models relevant for coastal and climate predictions and of learning biogeochemical dynamics—ultimately important not only for modeling the Earth system but also for understanding the world’s food system. In this latter Bayesian learning example, scientists can discover ecosystem reaction functions with Bayesian ML that assimilates observations in all of the possible candidate models, revealing the best model as well as the equations of a priori unknown model formulations, all with estimates of their uncertainties (Gupta et al., 2019; Haley et al., 2020; Lermusiaux et al., 2020; Lu and Lermusiaux, 2021). Next, Lermusiaux considered a deep learning example of neural closure models in which scientists may have some data and a simple model—for example, for climate prediction—and the goal is to add a closure rate to increase reliability without increasing computational cost (e.g., Gupta and Lermusiaux, 2021; Kulkarni et al., 2020; San and Maulik, 2018; Wan et al., 2018; Wang et al., 2020).
Turning from model optimization toward data optimization, Bayesian and ML adaptive sampling can be used to predict future observations that would optimally maximize information. Lermusiaux explained that such an approach could answer questions about where, when, and
what to observe to provide the most information; how much information is needed for ML methods; and how long such a system would remain predictable (Lermusiaux et al., 2017a,b). Example applications that he showed include optimizing where to release drifters—oceanographic floating devices that measure currents and other parameters—considering the reachability of the research ship releasing the drifters. Another example he presented is predicting the path of a tanker that would optimize energy use and safety, avoiding waves or bad weather.
Conveying Information and Uncertainty to End Users
Panelists discussed the challenges of conveying uncertainty to users and decision makers and outstanding gaps in making the information from ML/AI methods understandable. Rothenberg emphasized the importance of first understanding the end users and how they go about calculating risk and making decisions, and then mapping ML/AI capabilities to information that can be used to make effective decisions. Perhaps, for example, information about a weather nowcast could be distilled down to a binary “yes” or “no” reported with some confidence level. Cochran shared that for earthquake early warning, certain information has been removed from the product that ultimately gets used by the public because users only have seconds to take an action. Lermusiaux added that there are opportunities for funding agencies to connect social science researchers with Earth system science and modeling researchers to facilitate information exchange not only at the university and national laboratory levels but also at the high school level in efforts to educate the public on new concepts of Earth system modeling, data science, ML, probability, and uncertainty.
Panelists also discussed the range of stakeholder needs they consider in their work. At USGS, Cochran considers multiple products and users who all require different kinds of information: public-facing earthquake information, technical users including emergency managers and government officials, and private companies that may use USGS information to understand or reduce their earthquake hazard. From the perspective of the private sector and product management, Rothenberg shared that, in his experience, understanding end user priorities and decision-making processes is critical to conceptualize and ultimately deliver a data product or tool that might be used in a consumer context. He emphasized the importance of training a workforce that understands both the research and operations sides, both in Earth system science as well as ML/AI.
NOVEL FUNDING MECHANISMS, PARTNERSHIPS, AND KNOWLEDGE TRANSFER AMONG ACADEMIA, INDUSTRY, NONPROFITS, AND GOVERNMENT
Conversationalists shared their knowledge, ideas, and creative approaches to create mechanisms for funding opportunities, partnerships, and knowledge transfer among academia, industry, nonprofits, and government. In the context of opportunities and challenges for using ML/AI to advance Earth system science raised throughout the workshop—promising avenues for research, responsible and ethical use of AI, the urgency to diversify these fields, and strategies for training the workforce—several participants emphasized the need to develop sustainable mechanisms and funding sources to make progress in these areas. This final session brought
together panelists across sectors to discuss funding considerations and possibilities to enhance cross-disciplinary and cross-sector collaborations.
Funding Considerations to Advance ML/AI for Earth System Science
Participants discussed funding priorities over the next 5-10 years to make progress in ML/AI for Earth system science. Lynne Parker, White House Office of Science and Technology Policy, highlighted the importance of investing in frameworks that support large-scale interdisciplinary research. An example in the AI space, the National AI Research Institutes16 are funded by NSF and a number of other federal agencies and are inherently interdisciplinary, bringing together different academic fields and industry partners. Parker added that a network of institutes could add value, allowing institutes to share resources, best practices, and lessons learned.
Reflecting on the discussions around information exchange throughout the workshop, David Spergel, Simons Foundation, suggested that the movement of people and ideas between academia and industry would advance ML for Earth system science. Spergel noted that financial incentives and structures could encourage universities to embrace different career models—for example, an academic gaining real-world experience with ML applications in industry and then returning to academia with the insights and tools to train the next generation of scientists. Qingkai Kong, Lawrence Livermore National Laboratory, added that funding may also be needed to take ideas from academia or published in the scientific literature and develop and implement those tools to test them in real-world settings. Such mechanisms or incentives would facilitate knowledge transfer from the academic community toward products and the private sector.
Participants also reiterated discussions throughout the workshop about the opportunity to create an education track that would also allow Earth system scientists to train in ML/AI and open science. Kong shared his experience as an early career scientist trained in seismology and ML who could not find enough opportunities to marry the two communities in the past, and he suggested providing more opportunities for PhD students to explore career opportunities outside of academia.
Participants discussed the need for long-term support of software in the ML/AI and Earth system science space and for possibilities to advance its infrastructure. Chayes suggested building a dynamic infrastructure that scales to enable Earth system science—for example, built on an open science platform, such as a Jupyter, enabling transparency and collaboration. Spergel noted that the academic community often does not support careers that develop and maintain tools valuable to the community. Software support may also fall between the cracks at federal agencies where long-term software development is not incentivized or rewarded in the proposal process. Spergel suggested that there may be a role for science philanthropy to support this type of work. Gary Hattem, Independent Advisor, stressed that in order to attract investors who are interested in solving problems with a market return as well as philanthropists and high-net-worth individuals who are interested in interventions that benefit vulnerable communities, scientists should clearly define the constituency for the applications of their work and build trust with those communities that can advocate for new types of partnerships.
___________________
16 See https://beta.nsf.gov/funding/opportunities/national-artificial-intelligence-research-institutes
Leveraging Resources to Diversify ML/AI for Earth System Science
Participants identified the need to diversify the spaces that are tackling ML/AI and Earth system science issues. Chayes shared her experience at Berkeley where the data science major is growing faster than any other major and has allowed a wide cross-section of students to discover an interest and aptitude in data science by taking an entry-level course. In order to provide support and retain students in the program, Chayes gave the example of a data science scholars program that focuses on Pell Grant and first-generation students, offering a community of students extra help with course material. Spergel suggested providing a wider variety of on-ramps that would allow students to get expertise in data science in different ways—for example, fully funding master’s programs in data science or ML at minority-serving institutions, which would remove the cost barrier. Beyond providing grants, Spergel stressed the importance of developing strong programs and building capacity in these fields at Historically Black Colleges and Universities, Tribal Colleges and Universities, and other minority-serving institutions, and creating opportunities for experts in these domains to train the next generation.
Parker emphasized the importance of making infrastructure available to advance equity and diversity in AI. The National AI Research Resource Task Force17 was directed by Congress to design an implementation plan for a large-scale federated set of resources that would be widely accessible. The concept is to make resources available through a single tool with user and educational resources and provide access through a proposal system for data and computing resources. Parker explained that such a national resource would not only stimulate innovation but also diversify the students who can be trained and do research in the AI space.
Enhancing Multisector and Interdisciplinary Collaborations
Participants emphasized the importance of multisector meetings and platforms among academia, the private sector, and government to create positive feedback loops for interdisciplinary collaboration. Infrastructure to facilitate incentives and training could enable the ML/AI and Earth system science communities to come together to address societally relevant questions. Hattem suggested developing successful models of multisector and interdisciplinary collaborations that can be incrementally built on and socialized to introduce common languages and new ways of working. Furthermore, Hattem encouraged bringing people together who are mission-motivated through the development of new models that would harness the capabilities of local communities; organizing a social enterprise structure with continual impacts could garner investments from philanthropy and development banks. Stepping back, Spergel recognized that advances in ML/AI provide opportunities to make progress and gain deeper insights into climate and other pressing issues, particularly for students and professionals who are just starting their careers.
___________________
This page intentionally left blank.
