PLENARY SESSION: DEFINITIONS OF AND VISIONS FOR THE DIGITAL TWIN

During the plenary session, workshop participants heard presentations on the challenges and opportunities for Earth system digital twins, the history of climate modeling and paths toward traceable model hierarchies, and the use of exascale systems for atmospheric digital twins.

Umberto Modigliani, European Centre for Medium-Range Weather Forecasts (ECMWF), the plenary session's first speaker, provided an overview of the European Union's Destination Earth (DestinE) initiative,² which aims to create higher-resolution simulations of the Earth system that are based on models that are more realistic than those in the past; better ways to combine observed and simulated information from the Earth system; and interactive and configurable access to data, models, and workflows. More realistic simulations at the global scale could translate to information at the regional scale that better supports decision-making for climate adaptation and mitigation through tight integration and interaction with impact sector models. Now in the first phase (2021–2024) of its 7- to 10-year program, DestinE is beginning to coordinate with several other European initiatives, including Copernicus³ and the European Open Science Cloud.⁴

Modigliani explained that Earth system digital twins require unprecedented simulation capabilities––for example, ECMWF aspires to have a simulation on the order of 1–4 km at the global scale, which could enable the modeling of small scales of sea ice transformation and the development of more accurate forecasts. Earth system digital twins also demand increased access to observation capabilities, and efforts are under way to develop computing resources to leverage and integrate robust satellite information and other impact sector data. Furthermore, Earth system digital twins require exceptional digital technologies to address the opportunities and challenges associated with extreme-scale computing and big data. DestinE will have access to several pre-exascale systems via EuroHPC, although he pointed out that none of the available computing facilities are solely dedicated to climate change, climate extremes, or geophysical applications.

Modigliani indicated that improved data handling is also critical for the success of Earth system digital twins. The envisioned DestinE simulations could produce 1 PB of data per day, and he said that these data must be accessible to all DestinE users. While ECMWF will handle the modeling and the digital engine infrastructure for the digital twins, the European Organisation for the Exploitation of Meteorological Satellites will manage a data bridge to administer data access via a sophisticated application programming interface (for the computational work) and a data lake, which serves as a repository to store and process data that may be unstructured or structured. Policy makers could eventually access a platform operated by the European Space Agency to better understand how future events (e.g., heat waves) might affect the gross domestic product and to run adaptation scenarios. Current challenges include federating resource management across DestinE and existing infrastructures as well as collaborating across science, technology, and service programs. To be most successful, he emphasized that DestinE would benefit from international partnerships.

Venkatramani Balaji, Schmidt Futures, the plenary session's second speaker, presented a brief overview of the history of climate modeling, starting with a one-dimensional model response to carbon dioxide doubling in 1967. This early work revealed that studying climate change requires conducting simulations over long periods of time. He explained that as the general circulation model evolved in subsequent decades, concerns arose that its columns (where processes that cannot be resolved, such as clouds, reside) place restrictions on model structures, leading to slow progress. Parameterizing clouds is difficult because of their variety and interdependence. Clouds are also sensitive to small-scale dynamics and details of microphysics; parametrizing turbulence is necessary to understand how the small scale interacts with the large scale. He asserted that no resolution exists at which all features can be resolved unless extreme scales of computational capability are reached.

Balaji next discussed how model resolution has evolved over time, describing an example with only ~10 times improvement in 50 years. He suggested that climate models should be capable of 100 simulations of at least 100 years each in 100 days. Uncertainty—including chaotic (i.e., internal variability), scenario (i.e., dependent on human and policy actions), and structural (i.e., imperfect understanding of a system)—is also a key consideration for model design. Furthermore, he stressed that strong scaling is not possible with today's computers, which have become bigger but not faster.

Balaji noted that while various machine learning (ML) methods have been applied successfully for stationary problems (e.g., short forecasts), boundary conditions change over time in climate studies. One cannot use data for training into the future, because no observations of the future exist, he continued, but one could use short-running, high-resolution models to train simpler models to address the non-stationarity of climate. He observed that although digital twins were first introduced for well-understood engineered systems such as aircraft engines, digital twins are now discussed in the context of imperfectly understood systems. Even if climate and weather diverge further as model-free methods become successful in weather forecasting, he anticipated that theory will still play an important role. Model calibration is needed at any resolution, he added, as are methods for fast sampling of uncertainty. In closing, Balaji remarked that because high-resolution models should not play a dominant role, a hierarchy of models is needed. He suggested that ML offers systematic methods for calibration and emulation in a formal mathematical way to achieve traceable model hierarchies.

Mark Taylor, Sandia National Laboratories, the plenary session's final speaker, discussed the need for more credible cloud physics to achieve digital twins for Earth's atmosphere. For example, a global cloud-resolving model (GCRM) aims to simulate as much as possible using first principles. Although GCRMs are the “backbone” of any digital twin system, they require exascale resources to obtain necessary throughput and confront challenges in ingesting and processing results. He pointed out that computers are becoming more powerful but not necessarily more efficient; although GCRMs are expected to run well on exascale systems, they are expensive, with 1 megawatt-hour per simulated day required.

Taylor described his experience porting the following two types of cloud-resolving models to the Department of Energy's (DOE's) upcoming exascale computers:⁵ (1) SCREAM (Simple Cloud-Resolving Energy Exascale Earth System Model [E3SM]⁶ Atmosphere Model), the GCRM, runs at 3 km, with a goal to run at 1 km; and (2) E3SM-Multiscale Modeling Framework (MMF) is a low-resolution climate model with cloud-resolving “superparameterization.”

Taylor explained that SCREAM and E3SM-MMF are running well on graphics processing units (GPUs) systems. However, both use significant resources. GPU nodes consume 4–8 times more power than central processing unit (CPU) nodes but need to be 4–8 times faster. Furthermore, GPUs had only 5 times the performance on a per-watt basis in 10 years compared to CPUs. Until more powerful machines are developed, he anticipated that GCRMs will continue to run on GPUs, because GPUs are 1.5–3 times more efficient than CPUs on a per-watt basis.

PANEL 1: CURRENT METHODS AND PRACTICES

During the first panel, workshop participants heard brief presentations on methods and practices for the use of digital twins.

Yuyu Zhou, Iowa State University, discussed efforts to develop a model to reduce energy use for sustainable urban planning. The current model estimates energy use at the single building level and for an entire city of buildings by integrating the building prototypes, the assessor's parcel data, building footprint data, and building floor numbers, and it includes city-scale auto-calibration to improve the energy use estimate. The model enables the estimation of both the spatial pattern of energy use for each building and the hourly temporal pattern (e.g., the commercial area uses the least energy at midnight and the most at noon). He noted that the model could also be used to investigate the impacts of extreme events (e.g., heat waves) or human activities (e.g., building occupant behavior) on building energy use.

Zhou explained that real-time data from sensors and Internet of Things⁷ devices could be assimilated into the building energy use model to develop a digital twin with a more comprehensive, dynamic, and interactive visual representation of a city's buildings. Transportation data and social media data could also be integrated to enhance the representation of building occupant behavior. He asserted that a future digital twin could monitor real-time building energy use and optimize a city's energy performance in real time.

Christiane Jablonowski, University of Michigan, observed that the modeling community has created digital representations of reality for years; however, digital twins could integrate new capabilities such as high-resolution representations of climate system processes as well as advancements in ML and artificial intelligence (AI). She provided a brief overview of the current state of modeling and possible future trajectories, beginning with a description of typical phenomena at temporal and spatial scales (Figure 1). She said that the microscale will never be captured by atmospheric models in any detail, and the current generation of weather models (1–3 km grid scale) has reached its limit at the mesoscale. Modelers are currently most comfortable in the synoptic regime, the site of daily weather. Climate system timescales introduce new uncertainties, as they are in the boundary value problem category, not the initial value problem category. She added that the complexity and resolution of climate models has advanced greatly over the past several decades and will continue to increase with hybrid approaches and ML.

FIGURE 1

Hierarchy of temporal and spatial scales (left) and timeline of model complexities (right). SOURCE: Christiane Jablonowski, University of Michigan, presentation to the workshop. Left: Modified from The COMET Program (the source of this material is the (more...)

Jablonowski emphasized that selecting the appropriate spatial and temporal scales for digital twins is critical to determine the following: the phenomena that can be represented in a model; the correct equation set for the fluid flow; the required physical parameterizations, conservation principles, and exchange processes (i.e., a good weather model today is not a good climate model tomorrow); the model complexity (e.g., ocean, ice, and chemistry components are often not well tuned); decisions about coupling and related timescales that are key to making trustworthy predictions; and whether AI and ML methods as well as observations could inform and speed up current models.

Jean-François Lamarque, consultant and formerly of the National Center for Atmospheric Research's Climate and Global Dynamics Laboratory, shared the definition of a digital twin, as used in the digital twin Wikipedia entry: “a high-fidelity model of the system, which can be used to emulate the actual system … the digital twin concept consists of three distinct parts: the physical object or process and its physical environment, the digital representation of the object or process, and the communication channel between the physical and virtual representations.”⁸ He emphasized that a research question should determine the tool selection; in the current generation and in the foreseeable future, no single tool exists to answer all relevant questions. Thus, he suggested evaluating the advantages and disadvantages of each tool's approach as well as the timescales of interest. Three current tools include a coarse-resolution Earth system model, a high-resolution climate and weather model, and emulators and ML models. Key questions to evaluate the usefulness of each tool include the accuracy of the digital twin in representing the Earth system, the strength of the communication channel between the physical and digital representations, and the usefulness of the digital twin for climate research.

Gavin A. Schmidt, Goddard Institute for Space Studies of the National Aeronautics and Space Administration, explained that digital twins extend beyond the Earth system modeling that has occurred for decades. digital twins leverage previously untapped data streams, although whether such streams exist for climate change remains to be seen. Digital twins should explore a full range of possible scenarios, he continued, but only a small number of scenarios can be run with the current technology. A more efficient means to tap into information for downstream uses (e.g., urban planning) would be beneficial, and processes to update information regularly are needed. He added that higher resolution does not necessarily lead to more accurate climate predictions. Furthermore, ML does not overcome systematic biases or reveal missing processes.

Schmidt highlighted the value of improving the skill and usability of climate projections but noted that improvements in initial value problem skill do not automatically enhance boundary value problem skill. He stressed that no “digital twin for climate” exists, but digital twin technology could be used to strengthen climate models; for example, systematic biases could be reduced via ML-driven calibration, new physics-constrained parametrizations could be leveraged, and data usability could be improved.

PANEL 2: KEY TECHNICAL CHALLENGES AND OPPORTUNITIES

During the second panel, workshop participants heard brief presentations on challenges and opportunities for the use of digital twins.

Tapio Schneider, California Institute of Technology, provided an engineering-specific definition of a digital twin: a digital reproduction of a system that is accurate in detail and is updated in real time with data from the system, allowing for rapid experimentation and prototyping as well as experimental verification and validation where necessary. However, he pointed out that Earth system models are very different: achieving an accurately detailed digital reproduction of the Earth system is not feasible for the purposes of climate prediction, and the use of data for climate prediction (i.e., improving the representation of uncertain processes) is fundamentally different from the use of data for weather prediction (i.e., state estimation)—so continuously updating with data from the real system is less relevant for climate prediction. He suggested a three-pronged approach to advance climate modeling: (1) theory to promote parametric sparsity and generalizability out of observational data distributions, (2) computing to achieve the highest feasible resolution and to generate training data, and (3) calibration and uncertainty quantification by learning from computationally generated and observational data. Learning about processes from diverse data that do not come in the form of input–output pairs of uncertain processes is a key challenge, but he advised that new algorithms that accelerate data assimilation with ML emulators could address this problem.

Mike Pritchard, NVIDIA/University of California, Irvine, offered another description of a digital twin, which he suggested is a surrogate for a deterministic weather prediction system that, once trained, allows for much larger ensembles. Challenges exist in understanding the credibility of these tools as surrogates for the systems on which they are trained, and tests to ensure accountability would be beneficial. He noted that digital twins could help overcome the latency and compression barrier that prevents stakeholders from exploiting the full detail of the Coupled Model Intercomparison Project (CMIP) Phase 6⁹ library, and he described the potential value of pretraining digital twins to regenerate missing detail in between sparsely stored checkpoints at short intervals. Other challenges for the use of digital twins include the lack of available specialists to train ML models. Understanding the extrapolation boundary of current data-driven weather models is also important, as is finding ways to reproducibly and reliably achieve long-term stability despite inevitable imperfections. He concluded that digital twins offer a useful interface between predictions and the stakeholders who could leverage them.

Omar Ghattas, The University of Texas at Austin, described the continuous two-way flow of information between a physical system and a digital twin in the form of sensor, observational, and experimental data. These data are assimilated into the digital twin, and optimal decisions (i.e., control and experimental design) flow from the digital twin to the physical system. In such a tightly coupled system, if stable components are assembled incorrectly, an unstable procedure could emerge. He also explained that although real-time operation is not relevant to climate scales, it is relevant for climate-related issues (e.g., decisions about deploying firefighters to mitigate wildfires). Digital twins are built for high-consequence decision-making about critical systems, which demand uncertainty quantification (e.g., Bayesian data assimilation, stochastic optimal control), he continued, and real-time and uncertainty quantification settings demand reduced-order/surrogate models that are predictive over changing parameter, state, and decision spaces—all of which present massive challenges.

Abhinav Saxena, GE Research, depicted the three pillars of sustainability—environmental protection, economic viability, and social equity. Understanding how the environment and the climate are evolving in combination with how human behaviors are changing and how energy is being used impacts how critical infrastructure could be sustained. He noted that energy generation systems (e.g., gas and wind turbines) and other assets (e.g., engines that consume energy and produce gases) require detailed modeling to be operated more sustainably and efficiently; better weather and climate models are also needed to decarbonize the grid with carbon-free energy generation and microgrid optimization and to achieve energy efficient operations via energy optimization and resilient operation. He asserted that digital twins could help make decades-old systems that experience severe degradation and multiple anomalies more resilient. In this context of life-cycle sustainment, digital twins have the potential to guarantee reliability, optimize maintenance and operations, reduce waste and maximize part life, and reduce costs. He summarized that because physical engineering systems and Earth systems interact with each other, their corresponding digital twins should work in conjunction to best reflect the behaviors of these physical systems and subsequently optimize operations aided by these digital twins toward sustainability.

Elizabeth A. Barnes, Colorado State University, stated that because duplicating the complexity of the Earth system will never be possible, the term “digital twin” is misleading. With the explosion of ML, questions arise about the extent to which a digital twin would be composed of physical theory, numerical integration, and/ or ML approximations. Although achieving a “true digital twin” is unlikely, she explained that any future Earth system model will have information (e.g., observations) as an input leading to a target output (e.g., prediction, detection, discovery). She asserted that model developers should be clear about a model's intended purpose, and explainability (i.e., understanding the steps from input to output) is an essential component of that model's success. Explainability influences trust in predictions, which affects whether and how these tools could be used more broadly, allows for fine tuning and optimization, and promotes learning new science. She expressed her excitement about ML approaches that could integrate complex human behavior into models of the Earth system.

PANEL 3: TRANSLATION OF PROMISING PRACTICES TO OTHER FIELDS

During the third panel, workshop participants heard brief presentations from experts in polar climate, AI algorithms, ocean science, carbon cycle science, and applied mathematics; they discussed how digital twins could be useful in their research areas and where digital twins could have the greatest future impacts.

Cecilia Bitz, University of Washington, championed the use of digital twins to understand Earth system components such as sea ice. By increasing the realism of the ocean and atmosphere and achieving improvements in those components, improvements in downstream components could be realized. Highlighting opportunities for advances in sea ice components, she referenced ECMWF's high-resolution simulation of sea ice—with kilometer-scale flows of sea ice in the horizontal and large openings dynamically occurring—as an example of the progress enabled by moving to high resolution to view dynamic features. She expressed interest in broadening the kind of physics in the digital twin framework as well as developing parameterizations to be more efficient, which is not only a significant challenge but also a great opportunity to expand the capabilities of the surface component.

Anima Anandkumar, California Institute of Technology/ NVIDIA, discussed efforts to incorporate fine-scale features in climate simulations using neural operators. She explained that ML with standard frameworks captures only finite dimensional information; however, neural operators (which are designed to learn mappings between function spaces) enable finer scales, evaluate throughout the domain, and capture physics beyond the training data. Training and testing can be conducted at different resolutions, and constraints can be incorporated into the training. She encouraged using data-driven approaches to learn from a wealth of historical weather data and to incorporate extreme weather events into models—these approaches encompass a wide range of fields and new possibilities for generative AI in science and engineering.

Emanuele Di Lorenzo, Brown University, described work to engage coastal stakeholders and researchers to co-design strategies for coastal adaptation and resilience in Georgia. Supported by a community-driven effort, the Coastal Equity and Resilience (CEAR) Hub's¹⁰ initial modeling and forecasting capabilities provide water-level information at the scale where people live. A network of ~65 sensors that are distributed and interconnected wirelessly along the coast around critical infrastructure provides actionable data that are important for decision-makers, which are streamed into dashboards that are continuously redesigned with their input. He explained that as the focus on equity and climate justice expands, new factors related to resilience are emerging (e.g., health and socioeconomic well-being) that demand new community-driven metrics. The CEAR Hub plans to expand its sensor network to measure air and water quality and urban heating and combine these new sources of data with social data to develop the metrics.

Anna Michalak, Carnegie Institution for Science, explained that the carbon cycle science community focuses on questions around quantification (i.e., how greenhouse gas emissions and uptake vary geographically at different resolutions as well as their variability over time), attribution (i.e., constraining the processes that drive the variability seen in space and time, which requires a mechanistic-level understanding), and prediction (e.g., global system impact if emissions hold a particular trajectory or if climate changes in a particular way). These questions are difficult because carbon cycle scientists work in a data-poor environment—both in situ and remote sensing observations are sparse in space and time, and fluxes cannot be measured directly beyond the kilometer scale. The community has moved toward the use of ensembles of models as well as ensembles of ensembles to confront this problem. She said that the current best strategy to address the uncertainty associated with quantification, attribution, and prediction is to run multiple simulations of multiple models, with the whole community working on different incarnations of these models. She also suggested simplifying models before increasing their complexity to understand fundamental mechanistic relationships. The “holy grail” for digital twins in carbon cycle science, she continued, would be to use them for hypothesis testing, to diagnose extreme events, and for prediction. .

John Harlim, The Pennsylvania State University, defined digital twins as a combination of data-driven computation and modeling that could involve data simulation, modeling from first principles, and ML algorithms. He emphasized that the fundamental success of digital twins depends on their ability to compensate for the modeling error that causes incompatibility of numerical weather prediction models and climate prediction models. If one could compensate appropriately for this modeling error, the same model proven to be optimal for state estimation could also accurately predict climatological statistics. However, he asserted that this is not achievable for real-world applications. Specific domain knowledge is critical to narrow the hypothesis space of models, which would help the ML algorithm find the underlying mechanisms. Using a model-free approach, he continued, is analogous to allowing the algorithm to find a solution from a very large hypothesis space of models. Such a practice could reduce the bias, but this bias has to be balanced with the variance error. He stressed that the success of digital twins in other fields depends on whether enough informative training data exist to conduct reliable estimations.

PANEL 4: DISCUSSION ON TRANSPARENCY, SOCIETAL BENEFIT, AND EQUITY

Incorporating questions from workshop participants, Ye and Parris moderated the workshop's final discussion among three experts on transparency, societal benefit, and equity considerations for the use of digital twins: Amy McGovern, University of Oklahoma; Mike Goodchild, University of California, Santa Barbara; and Mark Asch, Université de Picardie Jules Verne. Ye invited the panelists to define digital twins. McGovern described digital twins as high-fidelity copies such that the digital twin and the original observations are as indistinguishable as possible and allow exploration of what-if scenarios. She added that a human dimension is crucial for the digital twin to enable decision-making. Focusing on the capacity for “high resolution,” Goodchild commented that a question arises about what threshold has been passed, as resolution has become finer and will continue to become finer. However, he noted that the human, decision-making, and sustainability components are defining characteristics of digital twins. Asch underscored the need to educate decision-makers in how to use the tools that scientists develop; much work remains to communicate that digital twins are decision-making tools, not “magic wands.” Parris inquired about how a digital twin extends beyond modeling. Goodchild highlighted the use of digital twins to visualize fine-resolution geospatial data, which will appeal to a broad audience, although visualizing uncertainty in these data is very difficult. McGovern explained that today's modeling world includes data in different formats and scales, with and without documentation—a digital twin could provide a consistent form to access these data, which could enable AI and visualization. Asch urged more attention toward improving the modeling of uncertainty, especially in light of recent advances in computational power. He emphasized that decisions derived from digital twins are probabilistic, based on the relationship between value and risk, not deterministic. In response to a question from Ye about unique approaches to visualize uncertainty, McGovern described an initiative where those creating visualizations are conducting interviews with end users to understand how uncertainty affects their trust in the model. A next step in the project is allowing the end users to manipulate the underlying data to better understand uncertainty.

Parris asked the panelists to share examples of digital twins that provide societal benefit. Goodchild described the late 1990s concept of the Digital Earth as a “prime mover” in this space and noted that the literature over the past 30 years includes many examples of interfaces between “digital twins” and the decision-making process, especially in industry. McGovern said that DestinE could allow people to explore how climate and weather will directly impact them; however, she stressed that more work remains for DestinE to reach its full potential. Asch suggested drawing from the social sciences, humanities, and political sciences to help quantify qualitative information. Integrating more diverse beliefs and values into science is critical, he continued, although enabling cross-disciplinary collaboration is difficult within existing funding streams. In response to a question from Parris about how digital twins integrate natural and human systems, Asch described work in the Philippines to model the spread of viral epidemics. He noted that creating dashboards is an effective way for end users to interact with a complicated problem; however, more work remains to model social and psychological phenomena. Goodchild highlighted the value of understanding interactions between humans and their environment in terms of attitudes and perceptions, and he referenced the National Science Foundation's coupled natural and human systems program and others' successes with agent-based modeling, where the behavior of humans is modeled through a set of rules.

Ye posed a question about the trade-offs of using data in a digital twin and maintaining privacy. Goodchild remarked that location privacy is becoming problematic in the United States. Location data are of commercial interest, which makes it even more difficult to impose regulations. He posited that the issue of buying and selling location data without individuals' awareness should be given more attention. With finer resolution, these ethical issues become critical for digital twins, and he suggested implementing a regulation similar to the Health Insurance Portability and Accountability Act (HIPAA). McGovern acknowledged the disadvantages of location data but also highlighted the advantages. She noted the need to evaluate trade-offs carefully to improve models while protecting privacy; a HIPAA-like regulation could make it difficult to obtain useful information for digital twins. Asch commented that the potential for abuse is significant, but high-level data-sharing agreements and data security technology could address this issue.

Parris inquired about the role of bias in digital twins. McGovern observed that although bias might be helpful, most often it is harmful and should be discussed more often in relation to weather and climate data. The first step is recognizing that bias exists at all stages of digital twin creation, and that systemic and historical biases affect which data are missing. Goodchild added that limiting the spatial extent of digital twins could help to address this issue, and Asch proposed that critical reasoning be used to detect bias. Given these issues, Ye asked how to improve confidence in digital twins. Asch stressed that transparency and reproducibility are key to increasing digital twin acceptance, and users should be able to follow a digital twin's reasoning as well as understand how to use and exploit it. McGovern stated that co-development with the end user helps advance both confidence and trustworthiness. Goodchild explained that some uncertainty in what digital twins predict and how they operate will always exist. He said that ethical issues related to digital twin reusability could also arise, and enforcing fitness for use is essential; “repurposing” is a significant problem for the software industry to confront.

Parris posed a question about strategies to build a diverse community of researchers for digital twins. Goodchild suggested first identifying the subsets of problems for which a digital twin could be useful. Asch said that understanding the principles of modeling and the problem-solving process for digital twins is key. He reiterated the value of bringing philosophy and science together to better develop and use these tools, emphasizing reasoning as a means to help democratize them. He also encouraged increasing engagement with students in developing countries. McGovern stressed that, instead of waiting for students to enter the pipeline, the current workforce should be given meaningful problems to solve as well as training on AI methods; furthermore, offering AI certificate programs at community colleges plays an important role in creating a more diverse workforce.

Suggested citation:

National Academies of Sciences, Engineering, and Medicine. 2023. Opportunities and Challenges for Digital Twins in Atmospheric and Climate Sciences: Proceedings of a Workshop—in Brief. Washington, DC: The National Academies Press. https://doi.org/10.17226/26921.