Challenges and Risks of Using ML/AI for Earth System Science
A number of challenges and risks are associated with the use of ML/AI, especially as these tools are applied to Earth system science. Issues of bias in ML/AI have been identified across domains, and Earth system science data, models, and problem selection all pose particular challenges with implications for diversity, equity, inclusion, justice, and ethics. Relatedly, as expertise at the intersection of ML/AI and Earth system science emerges, consideration may be required to equip the current workforce and the next generation of workers with new skill sets, and employers may need to build capacity to tackle emerging problems. In addition, as new ML/AI tools are developed, strategies may be needed to interface these new tools with existing Earth system science hardware, software, and data.
RESPONSIBLE AND ETHICAL USE AND DIVERSITY, EQUITY, INCLUSION, AND JUSTICE ISSUES FOR ML/AI IN EARTH SYSTEM SCIENCE
Panelists in this session considered what ethical standards should guide Earth system science, and how such standards relate to the lenses of diversity, equity, inclusion, and justice. Panelists also discussed potential biases in ML/AI approaches for Earth system science and promising ways to avoid those biases.
Ethical and Value Judgments Are Inherent in Earth System Science and the Deployment of ML/AI
David Danks, University of California, San Diego, began by challenging a common set of assumptions that ML/AI models—or computational systems and algorithms more generally—are “just math” or “just 1s and 0s,” which leads to the thinking that ethics and values matter only for the use of these models, while the algorithms themselves are objective and value-neutral. Danks explained that in fact, values are ubiquitous in ML/AI methods and there are many places throughout the pipeline of developing ML/AI systems where value choices are being made, even if they are not recognized by users that way.
Before building an ML/AI algorithm, values and ethical choices are made in how a problem is identified and specified. For example, decisions about the resolution of a model, what errors are acceptable, and what counts as success in an optimization problem are values choices that insert biases and preferences into a modeling system. After building an ML algorithm, there are nonpolicy values and ethics choices as well, including who has access to model outputs and descriptions and the reasons for making updates and revisions to a model. Danks emphasized that scientists are already engaged in ethics; he suggested that the scientific community recognize why various choices are being made, explain why they were defensible choices, and collaborate to improve those choices.
While ethical considerations have long been a part of Earth science in terms of problem identification and specification, panelists discussed how the use of ML/AI in Earth system science raises new and different challenges. Imme Ebert-Uphoff, Colorado State University, identified a concern that ethical decisions could be made by AI experts rather than the Earth system
scientists who have specific domain expertise. Abhishek Gupta, Montreal AI Ethics Institute, noted that the use of AI makes the scale and pace of impact for any application much larger. Priya Donti, Carnegie Mellon University, added that the use of ML/AI can accelerate existing systems and can create feedback loops that exacerbate inequalities and ethical failures, and the use of ML/AI tools may obscure or obfuscate important value discussions and divorce the models from important underlying value judgments and human decisions. Danks agreed and added that the use of ML/AI can introduce new opportunities to answer scientific questions in areas of Earth system science where difficult discussions about how to handle ethical issues have not yet been had.
Identifying and Addressing Biases in Data and ML/AI Methods
Ebert-Uphoff outlined the need for ethical, responsible, and trustworthy AI for environmental sciences. While AI can be a tool to advance environmental justice, the use of AI can also create new problems if scientists are not proactive (Box 3). Ebert-Uphoff provided examples of issues around biased training data that can lead to coded biases. Human-labeled data—for example, human-reported data of hail and tornadoes (Allen and Tippett, 2015; Potvin et al., 2019)—can create unintentional biases in which an ML algorithm would only predict hail and tornadoes in populated areas. Biases are also present in observations—for example, training data derived from sensors, because sensor networks often do not cover all populations equally, can be difficult to deploy in remote regions, and may only make measurements during daylight hours (Figure 7). Donti echoed the ways that datasets can be biased—for example, building
data that reflect the history of housing discrimination and redlining. There are also issues with AI algorithms themselves, including choices made during model development that can have a large impact on the result and AI algorithms learning undesired strategies (e.g., Lapuschkin et al., 2019), raising the question of who gets to make these decisions and how these decisions should be made. For example, in a task of detecting the locations of urban heat islands, the choice of spatial resolution would overlook small neighborhoods (low resolution) or introduce noise (high resolution), or sub-seasonal weather models may perform better in the mid-latitudes but not as well in other regions around the globe.
Panelists discussed possible strategies to address biases due to sparse spatial and temporal sampling in Earth system science data. Ebert-Uphoff suggested that the first step should be to come up with standard practices that can be applied to datasets to document and reveal the biases that exist but often go unrecognized. Relatedly, while open source and open science can help democratize AI, Danks noted that substantive metadata (e.g., the datasheet framework; see Gebru et al., 2020) will be required so that AI tools are not misinterpreted and misused. Donti added that the ML community is accustomed to using clean benchmark data to do algorithmic innovation and will require a culture shift to focus on questions around imperfect Earth science and climate change–related data. While existing methods already deployed in Earth sciences such as interpolation and imputation could be used to address sparse data problems, Danks explained that assumptions in those methods would also require thought so as not to introduce new biases. Domain-informed AI approaches, such as physics-informed ML, could embed nondata knowledge into a modeling system. Approaches from other domains such
as active learning in computer science—a form of ML in which an algorithm can choose which data it wants to learn from11—could use ML algorithms to identify locations where additional sampling or sensors would be beneficial. However, Danks cautioned that not every problem is solvable with ML/AI, and, in some cases, the answer may be that more instruments or sensors are needed to improve measurement capabilities.
Intersections among AI, Climate, and Sustainability
Donti discussed the intersection of AI, climate change, and values, and identified two ways to consider climate change and responsible AI: (1) employ AI in the context of climate action in a way that is fair, accountable, transparent, and equitable; and (2) make climate change a consideration for responsible AI by ensuring AI is broadly aligned with climate-relevant goals (Kaack et al., 2021). Donti broadened the discussion of bias to include three ways that AI can interact with or exacerbate issues of equity and justice: (1) large stakeholders in industrial agriculture have the resources to use AI methods to optimize a given problem, whereas smaller stakeholders may not have access to such resources, potentially exacerbating divides between industrial agriculture and smallholder farms; (2) ML/AI tools are readily available to entities with the most power in society and are often provided to a small set of powerful stakeholders, which could, for example, trap cities with costly urban analytics services (i.e., create a risk of “capture”); and (3) populations with fewer privacy regulations in place could be experimented on and taken advantage of.
Ebert-Uphoff cautioned that while XAI methods can be applied to weather applications, XAI is not sufficient to guarantee trustworthiness. Donti added that when employing AI for climate action, trustworthiness and accountability should be improved. For example, ensuring the safety and robustness of AI methods would be critical when optimizing electric power systems, and interpretability and auditability would be important in the context of high-stakes decisions about climate policy. Donti also explained how AI can be used in ways that directly facilitate climate change mitigation and adaptation strategies; however, AI can also be used in ways that accelerate or increase energy consumption and greenhouse gas emissions, and there is an energy cost associated with using computationally intensive AI models.
Gupta argued that scientists have an imperative to create sustainable AI systems, tying together responsible AI with the environmental impacts of deploying different technologies. Under the current paradigm, there has been a push toward larger AI models, which promote exploitative data practices, centralize power and homogeneity, and create a massive energy footprint. Gupta defined sustainable AI as a cohesive framework to mitigate the negative environmental and societal impacts of AI systems as they are designed, developed, and deployed. Several paths forward include elevating smaller models, creating alternative deployment strategies, increasing carbon efficiency and carbon awareness, and normalizing the accounting and reporting of carbon impacts of AI systems.
___________________
11https://deepai.org/machine-learning-glossary-and-terms/active-learning
Enforcing Ethics and Responsible AI
Panelists discussed how ethics and responsible AI could be enforced practically. Danks noted that market forces require a backdrop of standards, audits, and other mechanisms for transparency, including regulation and other frameworks. Donti agreed that policy and regulation have a role to play and that allowing tech-oriented companies and entities to act on good faith alone is not sufficient. Beyond public policy, organizations make many kinds of policy decisions about incentives, rules, and transparency requirements on small and large scales, and Donti shared the Global Partnership in AI12 multi-stakeholder effort that is making recommendations about how to embed values and ethics into the development of AI in the context of climate change (Figure 8).
Gupta provided examples of small-scale actions that can be taken now. ML/AI journals and conferences could require broader impact statements that would force scientists to consider the ethical impacts of the systems they are developing. Ebert-Uphoff added that journals and editors have an important role to play in setting standards for disclosure and transparency in algorithms and value decisions as part of the article submission and acceptance process.
Linking Communities
Danks shared examples from the design and technology design communities for ensuring that values and interests of made-vulnerable populations are represented in ML/AI for
___________________
12 See https://gpai.ai
Earth system science. These approaches include community-supported and community-engaged design processes. One of the lessons from these approaches is that community-engaged design is an iterative, cyclical process that never ends because scientists are continually learning from communities about their values and reinforcing those values through science and technology. Gupta agreed that meaningfully involving stakeholders is an outstanding challenge that is essential to pursue.
Donti emphasized that responsible AI in the Earth system and climate change–related context would require bringing together different research perspectives including stakeholders’ lived experiences, respecting bodies of knowledge that already exist in these domains, and respecting different ways of thinking. Danks expanded the idea of domain-informed and domain-educated ML/AI researchers, suggesting that Earth scientists collaborate with not only AI researchers but also ethicists and sociologists. Many domains and sectors are tackling these same challenges and could benefit from building collaborations and integrating the perspectives, frameworks, and techniques of different disciplines.
WORKFORCE DEVELOPMENT CAPACITY AND SKILL SETS
Conversationalists in this session discussed gaps in education for those working at the intersection of ML/AI and Earth system science, needs and strategies for the private sector and academia in workforce development, the role of continued education for the current workforce, and capacity building at earlier educational stages.
Gaps in Training and Opportunities for Continuing Education
The panelists began by discussing gaps in education among students just entering the workforce. Hamed Alemohammad, Radiant Earth Foundation, echoed the gap discussed throughout the workshop between domain scientists who have only used ML tools “off the shelf” and computer scientists who are not knowledgeable about the Earth science domain, forcing employers to choose between the two in the hiring process. Terri Adams, Howard University, explained that this gap points to a need for interdisciplinary education and training, and Lak Lakshmanan, Google, added that competency with industrial-scale datasets is also an educational gap that should be filled. Building interdisciplinary educational programs offers an opportunity to think about how ML can be adapted for Earth system science. Professional societies could also play a role in convening the Earth system science community to identify educational needs on a 5- to 10-year time horizon.
Lakshmanan noted that computer science courses are often more project-driven than those in the Earth sciences, and the Earth sciences could require code submission and verification or forecasting competitions to build skills more common in data science fields. Alemohammad agreed that teaching coding and programming as part of an Earth science curriculum is key; research and production level code could be brought closer together in the classroom by teaching reproducible coding, for example, and software development could be incorporated into proposals so that there would be resources allocated toward learning those skills. Adams noted the challenges of only being able to spend a limited time with students in the classroom over a semester or quarter, as well as the importance of continual dialogue with the private sector to understand its needs. Rebecca Nugent, Carnegie Mellon University,
encouraged educators to incorporate ML/AI and data science concepts into curricula early, utilize known pedagogical tools, and document and share best practices with the community.
Nugent shared that many companies in private industry are asking universities to design professional development programs for employees at every level in their workforces. Panelists discussed existing continuing education resources as both experiential learning programs that are part of learning communities, as well as online courses or certificates that can be done independently (Box 4). Alemohammad noted that many of these courses are restricted to computer science; there is a need for the development of interdisciplinary courses that focus on adapting ML/AI for Earth system science applications rather than adopting them.
Sparking Interest from Early Ages
Adams shared an example of summer camps for high school students interested in weather or climate through the NOAA Cooperative Science Center for Atmospheric Sciences and Meteorology. Adams saw those students pursue degrees in atmospheric science or meteorology in college or graduate school, and suggested that this may be a good model to expose students
to fields they may not have otherwise considered. Nugent added that beyond coding camps that have become more prevalent, data literacy—including collecting and organizing data and drawing conclusions—should be incorporated into education starting at an early age, and activities could put these ideas in the context of Earth system science. Adams agreed that connecting younger learners with applications in their everyday lives (e.g., how AI is used in their smart home devices) could help to cultivate their interest in these subjects.
Cultivating Diversity in the Workforce and Strategies to Build Capacity
Panelists began by acknowledging the concept of the leaky pipeline at all stages of the career track and the importance of providing educational and career opportunities to a diverse population. Panelists discussed the need for continued investment in programs that have demonstrated success—for example, NOAA Cooperative Science Centers include workforce development as part of their core mission, and NSF’s Research Experience for Undergraduates (REU) program has expanded access to research experiences for students at smaller undergraduate institutions. Nugent challenged the community to expand the reach of REU programs to benefit broader populations by providing more accessibility, for example, through virtual research opportunities. Alemohammad suggested that private employers should think about how they can help provide opportunities at the university level for underrepresented minorities so that the applicant pool is diversified by the time they reach the hiring stage. Panelists agreed that one concrete step would be to ensure that internships are paid in order to make these opportunities available to students beyond those who can do unpaid work.
CHALLENGES AND OPPORTUNITIES FOR EARTH SCIENCE TECHNOLOGY AND DATA
During this session, panelists considered open data, standards, and platforms to facilitate open science for ML/AI and Earth system science as well as technology development, funding models, and education challenges and opportunities for Earth system science technology and data. Panelists were asked to consider the following questions: What have open data enabled for ML/AI that would otherwise not be feasible? What are the main stumbling blocks to advance open science data in support of ML/AI applications? What are potential top priorities to tackle over the next 5–10 years with respect to Earth system science technology and data? In this panel, the current realities of public and private organizations each building their own ML/AI for Earth system science “data platforms” met visions of an open, inclusive, and integrated platform for big-data Earth system science.
Challenges of the Current Paradigm
Ryan Abernathey, Columbia University, observed that the data-intensive nature of ML/AI research has forced the scientific community to confront the limitations of current data infrastructure and software in Earth system science. While the volume of Earth science datasets is expected to grow dramatically, the current data infrastructure assumes scientists will download and analyze data on their local computers (Figure 9). Ana Pinheiro Privette, Amazon Sustainability Data Initiative (ASDI), agreed that because Earth science data are highly
distributed, every user who wants to apply ML/AI to Earth system science is required to identify the critical datasets and acquire, store, and clean the data—all before using the data—which accounts for about 80 percent of a user’s effort (Patil and Mason, 2014). Katie Dagon, National Center for Atmospheric Research (NCAR), reiterated how time intensive it currently is to identify and prepare data, the need to share and reuse parts of the data analysis work flow, and how confronting these challenges will enable the research community to better apply ML/AI tools for modeling and analysis. Abernathey argued that the current framework also limits ambitious ML/AI research and inclusion within the geosciences because only privileged research institutions have the capabilities to do research at the scale required. ML/AI tools are exacerbating existing infrastructure challenges because researchers now want to download entire datasets to train their ML models.
Jason Hickey, Google, outlined trends in Earth science technology including sustained investment in Earth observing systems and environmental data systems—satellites, ground stations, airplanes, vehicles, buoys, etc. While there has been a large growth in the amount of data being collected, these data are not being used; for example, Hickey stated that only 3-5 percent of satellite data are being used in weather forecasting. ML/AI tools could be critical for managing these massive amounts of data.
At the same time, Hickey recognized the growth of scalable cloud platforms that are globally accessible and generally publicly available. Chelle Gentemann, Farallon Institute, shared how advances in technology have transformed her workflows such that she can access a coding environment through a web browser on an internet-connected device and connect quickly to a massive server farm, petabytes of data, and other hubs (Figure 10). However, she noted that the real goal would be to bring together groups of people with different expertise to work together to increase the pace and expand the impact of the science.
Promising Paths Forward
Each of the panelists shared their own examples of ways to build open science platforms and infrastructure to overcome barriers of the current tools and systems. Abernathey shared Pangeo13 as an example of an open science data community that is prototyping a way of working
___________________
13 See https://pangeo.io
for building data-intensive software and infrastructure. Pangeo centers on scientific users and use cases, contributes to existing open source software, and deploys that infrastructure in various cloud computing and high-performance computing environments.
Federal agencies are also investing in the ML/AI and data space and integrating open science principles. For example, NASA’s Transform to OPen Science (TOPS)14 program is a 5-year, $40 million investment to increase understanding and adoption of open science standards and techniques. Gentemann explained that the objectives of the program are to increase understanding and adoption of open science, accelerate major scientific discoveries, and broaden participation by underrepresented communities.
Privette shared the example of how ASDI is using cloud computing to democratize the ML/AI space by significantly reducing the cost, time, and technical barriers associated with analyzing large datasets to generate sustainability insights. ASDI is identifying critical datasets needed to solve climate and sustainability challenges, subsidizing the data storage, making the data easily accessible on Amazon’s cloud platform, and enabling users to experiment by offsetting the cost of experimentation through dedicated cloud grants.
At NCAR, Earth System Data Science (ESDS),15 an interdisciplinary grassroots project, formed with the vision to increase the effectiveness of the NCAR workforce by promoting deeper collaboration centered on analytics and big data, which would allow scientists to serve the
___________________
university community and deliver actionable, reproducible science. Dagon explained that, practically, this effort has required building a community and overcoming cultural, communication, and scientific challenges. The ESDS approach has been to cultivate a community of practice focusing on core software development and computing, training and education, and communication and best practices.
Longer-term Visions for the Future of Open Science
Imagining the landscape in 20 years, Hickey suggested that open, collaborative science will allow teams to share data and models efficiently. In terms of environmental information delivery, Hickey suggested imagining a more personalized and immersive experience where, for example, digital twins—models of the Earth as it is—can be used in a massive multiplayer online game for play and information delivery. Hickey emphasized the importance of fostering environmental justice and making information open to all communities so that underserved communities have equal access to information.
Abernathey shared a vision for an extensible, interconnected federation of high-performance Earth system data analytics environments to which scientists and institutions could bring capacity, data, and computing (Abernathey et al., 2021; de La Beaujardière, 2019; Gentemann et al., 2021). Abernathey also emphasized that realizing the potential of ML/AI in Earth system science will require prioritizing data and software engineering and development as much as the private sector does (Sambasivan et al., 2021). Additionally, Dagon outlined how journals, funding agencies, and performance evaluations can play a role in balancing the culture of static scientific publishing with the dynamic expectations of open science and development by creating incentives.
Opportunities to Shift Culture
Participants discussed existing barriers and strategies to overcome them to realize the open science future imagined by the panelists. Given the significant computing skills that ML/AI require, Dagon suggested creating and sharing analysis workflows for ML tasks using a set of best practices and examples, as well as developing benchmark datasets to facilitate wider application of ML/AI tools. Abernathey challenged the community to reject “not invented here” syndrome—the tendency to discount existing solutions that do not strictly meet the needs of a new project—and collaborate to extend existing open source software to focus on delivering on community needs rather than branding a product as one’s own. Gentemann identified a need to motivate substantial contributions to open source infrastructure that is underlying so much science; the involvement of federal agencies and their workforce could be a powerful way to stabilize and advance open source infrastructure. Hickey added from the private sector perspective: For-profit organizations like Google actively promote the development and information sharing of their open source community without sacrificing their for-profit motive in the process.
Panelists acknowledged the importance of making the open science community more diverse and inclusive. Gentemann added that as the community transitions from local infrastructure to cloud computing, care and strategy will be required to avoid replicating existing inequalities. In order to realize a vision of an open science platform accessible to all, Hickey recognized that there will be a cost associated with using the platform, and equitable science will
require subsidies to empower science in underfunded countries. Panelists agreed that cloud access and cloud storage and computing credits are not enough for equitable access; developing a global infrastructure and building capacity to help communities use these resources will also be critical.
Relatedly, with respect to the large amount of ML/AI courses and trainings available, Dagon identified the challenge to provide accessibility and entry points for both domain scientists and ML/AI experts. Abernathey suggested offering data engineering courses and degrees and working with industry to identify needed skills and training. Hickey added that the cloud—an open science platform—also enables education about the use of open science that can become globally available.
