Closing Thoughts
The workshop underscored broad interest in and unique challenges for using ML/AI approaches to advance Earth system science and its applications. At the same time, the dialogue among Earth system science experts, ML/AI researchers, social and behavioral scientists, ethicists, and decision makers from academia, government, and the private sector facilitated by this workshop also surfaced challenges and risks ranging from data and technology issues to ethical and workforce considerations. Overall, participants recognized opportunities for the convergence of ML/AI approaches with the Earth system science field to facilitate knowledge discovery; improve the usability and accessibility of existing Earth science data workflows; and address the information needs of decision makers, communities, and other stakeholders.
EMERGING ML/AI APPROACHES AND OPPORTUNITIES FOR MULTIDISCIPLINARY COLLABORATION
Earth system science presents enormous computing and analysis challenges. Using ML/AI tools offer opportunities to improve predictive models, utilize computing resources in a way that is faster and less costly, improve understanding of the Earth system, and bring together disparate fields of study. For Earth system science applications, participants discussed the importance of benchmark datasets, which allow a quantitative evaluation of ML approaches and a separation of concerns among domain scientists, ML experts, and high-performance computing experts. Participants also emphasized the importance of embedding physical laws and prior knowledge within ML algorithms, especially in fields such as climate science where researchers seek to detect small signals in a noisy system and regimes may shift; these considerations are not typically accounted for in conventional ML approaches. Participants also discussed how ML/AI could be incorporated into the traditional data assimilation framework used in Earth system science.
ML/AI tools could also be used to understand and integrate social and human engineered systems into Earth system science. Participants noted that these types of applications would be highly relevant to advancing climate change decision support specifically and could help illuminate the socioeconomic asymmetries that will be exacerbated by climate change. Key gaps were identified, such as the flow of uncertainty, causality and linkages between the different parts of the complex Earth system, the ultimate connection between these physical models and risk, and how to account for uncertainties in human behavior.
As Earth system scientists increasingly incorporate ML/AI methods into their work, several participants emphasized that XAI—insights into where decisions are being made in AI algorithms—and interpretable AI—understanding of how and why decisions are made by an algorithm—are approaches that offer opportunities to uncover problematic strategies, build trust, choose appropriate tools or algorithms, and learn something new. Several workshop participants suggested that interpretability of an ML/AI method should be a design goal from the outset of a workflow in order to advance science.
Several participants noted that a central challenge in using ML/AI to advance Earth system science is bridging the gap between domain scientists and ML/AI experts in the philosophy, vocabulary, practices, purposes, and benchmarks of the two communities. Off-the--
shelf ML tools are often not sufficient for Earth system applications. Thus, engaging domain scientists in the design of ML/AI approaches from the beginning could bring physical and prior scientific knowledge into the ML/AI design. Participants identified an opportunity to develop and foster “ML domain scientists” who would have expertise in both of these fields.
FACILITATING OPEN SCIENCE
Advances in data, software, and computing are enabling transformational changes in the way science is done. Participants shared visions of scientists with different areas of expertise working together to increase the pace and impact of the science being done at the intersection of Earth system science and ML/AI. Current workflows for data analysis—including highly distributed data, time intensive data wrangling, and limited sharing and reuse of the data analysis workflow—are inefficient and present barriers to collaboration, inclusion, and knowledge transfer. The data-intensive nature of ML/AI research exacerbates these problems, forcing the Earth system science community to confront the limitations of the current paradigm. In recent years, there has been a sustained investment in Earth observing systems and models, as well as the growth of global, scalable cloud storage, computing, and infrastructure. Participants explained that these capabilities are not yet being utilized effectively, and open science provides an opportunity to support global creativity and a more broadly inclusive research community. However, making data open is not the same as fully democratizing science—participants identified other steps that could be useful, including preparing data for analysis, offsetting the cost of experimentation, and bringing communities together to address blind spots and biases to cultivate an inclusive community of practice.
In order to achieve open data, open source, and open knowledge, many participants identified the need for a culture shift that rejects the “not invented here” attitude and encourages inclusion and collective contributions to community code. Much of the data and computing tools are publicly generated, and both the public and private sector would likely benefit from a common unified infrastructure that supports multisector collaborations. Participants also suggested building capacity and making subsidies for resources like cloud computing available to enable equitable access and avoid replicating existing inequalities.
RESPONSIBLE AND ETHICAL USE OF ML/AI APPROACHES
Workshop participants highlighted the importance of ethical and responsible use of ML/AI in Earth system science. Values and ethical choices are already being made before and after ML/AI tools are introduced, and participants discussed the importance of recognizing, explaining, and improving these choices. Specifically for Earth system science, there are ethical challenges with training data, ML/AI models, workforce training, and interactions with society. Several participants identified a need for sustainable AI to address challenges with the current paradigm (e.g., exploitative data practices, barriers to entry, energy footprint) as well as opportunities to integrate responsible AI into goals for climate action. Ethical questions at the intersection of ML/AI and Earth system science pose unique and urgent challenges: The scale and pace of impact of ML/AI approaches may be larger than existing Earth system science tools; ethical decisions about Earth system science issues may be made by nonexperts; ML/AI have the
capacity to create feedback loops and exacerbate disparities; and ML/AI could open new science questions for which ethical dimensions may not have been considered.
EDUCATING THE WORKFORCE AT THE INTERSECTION OF ML/AI AND EARTH SYSTEM SCIENCE
Several educational challenges and opportunities at the intersection of Earth system science and ML/AI were raised throughout the workshop. Participants identified gaps in education for those entering the workforce at the intersection of ML/AI and Earth system science. Employers are looking for production level skills and workers who know how to optimize, automate, and process large amounts of data. Those entering the workforce would likely benefit from training on tools that scale to real-world datasets. A vision was suggested in which data science would be fully embraced as part of a domain science curriculum in order to ensure that ML/AI are properly leveraged. Such a curriculum could include, for example, requiring learning the fundamentals of data analysis, statistics, and scientific computing as a basis for launching into an ML/AI application, as well as developing a climate or Earth system data science discipline. Software development and data science could be incorporated as part of proposals and projects instead of treated as an afterthought, and experiential learning through partnerships with industry could be valuable. While many online courses and tutorials are available in ML/AI, there is a lack of Earth science applications featured in these resources, and building learning communities to develop the workforce was suggested.
An urgent need was also identified by many participants to broaden the participation of minoritized groups given the lack of diversity and equity in the geosciences and computer sciences. In order to prioritize diversity, equity, and inclusion in the workforce, participants discussed the need for continued investment and the creation of pathways. While many strategies would be needed to create pathways into the workforce, examples may include paid internships; creative teaching approaches and curricula; and partnerships with industry to adopt, adapt, and improve strategies. Participants discussed potential roles for federal agencies and professional organizations to address workforce development challenges, strengthen partnerships between academia and the private sector, and expand the accessibility of existing programs like NSF’s REUs.
USING ML/AI FOR DATA-DRIVEN DECISION MAKING
One application of Earth system science is real-time decision making, and participants discussed the role of ML/AI in bringing together physical models, prior knowledge, and data-driven mechanisms to serve the needs of end users. Rapidly growing large and structured datasets have enabled ML/AI to be used effectively, particularly when coupled with physics-based models. ML could be an important tool to produce probabilistic forecasts, including uncertainty quantification about the likelihood of a forecast, which can inform decision making. Many challenges in this space remain, including advancing from success with simple problems to more complex problems, breaking down barriers to entry for learning ML/AI methods, and considering the fragility and robustness of ML/AI when making decisions. Participants also discussed how to meet the needs of the end user—for example, how much information to provide and which decision metrics to use—and the remaining challenges of translating research advances into operational practice.
FUNDING OPPORTUNITIES TO ACCELERATE PROGRESS
Participants discussed funding considerations to make progress in research, applications, and education at the intersection of ML/AI and Earth system science. Mechanisms may be needed to facilitate large-scale interdisciplinary work through novel partnerships with government agencies, academia, private industry, and nonprofits. The expansion of interdisciplinary homes for young scholars may require incentives to break down disciplinary silos in academia. Diversifying the fields of ML/AI and Earth system science was a repeated theme in the funding context—particularly providing a wider variety of on-ramps and offsetting the costs of training or retraining students. The role of private foundations was described as playing a different but complementary role with government in order to improve the overall operational structure. In addition, participants highlighted opportunities for philanthropists and development banks to fund projects with social impacts.
***
Enthusiastic participation of both the Earth system science and ML/AI communities in the workshop indicates wide interest in fostering collaborations and knowledge exchange between these domains. Participants anticipated that ML/AI methods will ultimately be tools that will be used in combination with, for example, inverse methods and data assimilations, rather than a magic bullet to solve every Earth system science problem. Participants offered actionable steps to make progress, including creating educational pathways so that students and professionals who understand Earth system science can innovate in ML/AI, incorporating ethical and responsible AI into Earth system science applications, institutionalizing the use of open science and open data, pursuing research in how people make decisions, and developing infrastructure to support these innovations. The intersection of ML/AI and Earth system science is ripe for creative thinking around partnerships, knowledge transfer, and funding mechanisms to support advancements in societally relevant science.
