2

The Digital Twin Landscape

This chapter lays the foundation for an understanding of the landscape of digital twins and the need for an integrated research agenda. The chapter begins by defining a digital twin. It then articulates the elements of the digital twin ecosystem, discussing how a digital twin is more than just a simulation and emphasizing the bidirectional interplay between a virtual representation and its physical counterpart. The chapter discusses the critical role of verification, validation, and uncertainty quantification (VVUQ) in digital twins, as well as the importance of ethics, privacy, data governance, and security. The chapter concludes with a brief assessment of the state of the art and articulates the importance of an integrated research agenda to realize the potential of digital twins across science, technology, and society.

DEFINITIONS

Noting that the scope of this study is on identifying foundational research gaps and opportunities for digital twins, it is important to have a shared understanding of the definition of a digital twin. For the purposes of this report, the committee uses the following definition of a digital twin:

This definition is based heavily on a definition published in 2020 by the American Institute of Aeronautics and Astronautics (AIAA) Digital Engineering Integration Committee (2020). The study committee’s definition modifies the AIAA definition to better align with domains beyond aerospace engineering. In place of the term “asset,” the committee refers to “a natural, engineered, or social system (or system-of-systems)” to describe digital twins of physical systems in the broadest sense possible, including the engineered world, natural phenomena, biological entities, and social systems. The term “system-of-systems” acknowledges that many digital twin use cases involve virtual representations of complex systems that are themselves a collection of multiple coupled systems. This definition also introduces the phrase “has a predictive capability” to emphasize the important point that a digital twin must be able to issue predictions beyond the available data in order to drive decisions that realize value. Finally, the committee’s definition adds the sentence “The bidirectional interaction between the virtual and the physical is central to the digital twin.” As described below, the bidirectional interaction comprises feedback flows of information from the physical system to the virtual representation and from the virtual back to the physical system to enable decision-making, either automatic or with a human- or humans-in-the-loop. Although the importance of the bidirectional interaction is implicit in the earlier part of the definition, our committee’s information gathering revealed the importance of explicitly emphasizing this aspect (Ghattas 2023; Girolami 2022; Wells 2022).

While it is important to have a shared understanding of the definition of a digital twin, it is also important to recognize that the broad nature of the digital twin concept will lead to differences in digital twin elements across different domains, and even in different use cases within a particular domain. Thus, while the committee adopts this definition for the purposes of this report, it recognizes the value in alternate definitions in other settings.

Digital Twin Origins

While the concept itself is older, the term “digital twin” emerged around 2010 during technical roadmapping efforts at NASA co-led by John Vickers. The term “digital twin” was defined in published NASA reports by Piascik et al. (2012) and Shafto et al. (2012), and in a follow-on paper by Glaessgen and Stargel (2012):¹

___________________

¹ This paragraph was changed after the release of the report to accurately reflect the emergence of the term “digital twin.” These NASA reports were released publicly in 2010 but have 2012 official publication dates.

This definition and notion are built on earlier work by Grieves (2005a,b) in product life-cycle² management. A closely related concept is that of Dynamic Data Driven Application Systems (DDDAS) (Darema 2004). Some of the early published DDDAS work has all the elements of a digital twin, including the physical, the virtual, and the two-way interaction via a feedback loop. Many of the notions underlying digital twins also have a long history in other fields, such as model predictive control, which similarly combines models and data in a bidirectional feedback loop (Rawlings et al. 2017), and data assimilation, which has long been used in the field of weather forecasting to combine multiple sources of data with numerical models (Reichle 2008).

Much of the early work and development of digital twins was carried out in the field of aerospace engineering, particularly in the use of digital twins for structural health monitoring and predictive maintenance of airframes and aircraft engines (Tuegel et al. 2011). Today, interest in and development of digital twins has expanded well beyond aerospace engineering to include many different application areas across science, technology, and society. With that expansion has come a broadening in the views of what constitutes a digital twin along with differing specific digital twin definitions within different application contexts. During information-gathering sessions, the committee heard multiple different definitions of digital twins. The various definitions have some common elements, but even these common elements are not necessarily aligned across communities, reflecting the different nature of digital twins in different application settings. The committee also heard from multiple briefers that the “Digital Twin has no common agreed definition” (Girolami 2022; NASEM 2023a,b,c).

ELEMENTS OF THE DIGITAL TWIN ECOSYSTEM

A Digital Twin Is More Than Just Simulation and Modeling

The notion of a digital twin builds on a long history of modeling and simulation of complex systems but goes beyond simulation to include tighter integration between models, observational data, and decisions. The dynamic, bidirectional interaction between the physical and the virtual enables the digital twin to be tailored to a particular physical counterpart and to evolve as the physical counterpart evolves. This, in turn, enables dynamic data-driven decision-making.

Finding 2-1: A digital twin is more than just simulation and modeling.

Conclusion 2-1: The key elements that comprise a digital twin include (1) modeling and simulation to create a virtual representation of a physical

___________________

² For the purposes of this report, the committee defines life cycle as the “overall process of developing, implementing, and retiring … systems through a multistep process from initiation, analysis, design, implementation, and maintenance to disposal” as defined in NIST (2009).

counterpart, and (2) a bidirectional interaction between the virtual and the physical. This bidirectional interaction forms a feedback loop that comprises dynamic data-driven model updating (e.g., sensor fusion, inversion, data assimilation) and optimal decision-making (e.g., control, sensor steering).

These elements are depicted abstractly in Figure 2-1 and with examples in Box 2-1. More details are provided in the following subsections.

The Physical Counterpart and Its Virtual Representation

There are numerous and diverse examples of physical counterparts for which digital twins are recognized as bringing high potential value, including aircraft, body organs, cancer tumors, cities, civil infrastructure, coastal areas, farms, forests, global atmosphere, hospital operations, ice sheets, nuclear reactors, patients, and many more. These examples illustrate the broad potential scope of a digital twin, which may bring value at multiple levels of subsystem and system modeling. For example, digital twins at the levels of a cancer tumor, a body organ, and a patient all have utility and highlight the potential trade-offs in digital twin scope versus complexity. Essential to being able to create digital twins is the ability to acquire data from the physical counterpart. These data may be acquired from onboard or in situ sensors, remote sensing, automated and visual inspections, operational logs, imaging, and more. The committee considers these sensing and observational systems to be a part of the physical counterpart in its representation of the digital twin ecosystem.

The virtual representation of the physical counterpart comprises a computational model or set of coupled models. These models are typically computational representations of first-principles, mechanistic, and/or empirical models, which take on a range of mathematical forms, including dynamical systems, differential equations, and statistical models (including machine learning [ML] models). The set of models comprising the virtual representation of a digital twin of a complex system will span multiple disciplines and multiple temporal and spatial scales.

Digital twin examples in the literature employ models that span a range of fidelities and resolutions, from high-resolution, high-fidelity replicas to simplified surrogate models.

Another part of the digital twin virtual representation is the definition of parameters, states, and quantities of interest. The computational models are characterized by parameters that are the virtual representation of attributes such as geometry and constitutive properties of the physical counterpart, boundary

conditions, initial conditions, external factors that influence the physical counterpart, and transfer coefficients between resolved processes and parameterized unresolved processes. Sometimes these parameters will be known, while in other cases they must be estimated from data. Some types of models may also be characterized by parameters and hyperparameters that represent numerical approximations within a model, such as Gaussian process correlation lengths, regularization hyperparameters, and neural network training weights. The committee refers to this latter class of parameters as numerical model parameters to distinguish them from the parameters that represent attributes of the physical system. The committee uses the term state to denote the solved-for quantities in a model that takes the form of a dynamical system or system of differential equations. However, the committee notes that in many cases, the distinction between parameter and state can become blurred—when a digital twin couples multiple models across different disciplines, the state of one model may be a parameter in another model. Furthermore, the committee notes that many digital twin use cases explicitly target situations where parameters are dynamically changing, requiring dynamic estimation and updating of parameters, akin to state estimation in classical settings. Lastly, the committee denotes quantities of interest as the metrics that are of particular relevance to digital twin predictions and decisions. These quantities of interest are typically functions of parameters and states. The quantities of interest may themselves vary in definition as a particular digital twin is used in different decision-making scenarios over time.

An important theme that runs throughout this report is the notion that the virtual representation be fit for purpose, meaning that the virtual representation—model types, fidelity, resolution, parameterization, and quantities of interest—be chosen, and in many cases dynamically adapted, to fit the particular decision task and computational constraints at hand. Another important theme that runs throughout this report is the critical need for uncertainty quantification to be an integral part of digital twin formulations. If this need is addressed by, for example, the use of Bayesian formulations, then the formulation of the virtual representation must also define prior information for parameters, numerical model parameters, and states.

Bidirectional Feedback Flow Between Physical and Virtual

The bidirectional interaction between the virtual representation and the physical counterpart forms an integral part of the digital twin. This interaction is sometimes characterized as a feedback loop, where data from the physical counterpart are used to update the virtual models, and, in turn, the virtual models are used to drive changes in the physical system. This feedback loop may occur in real time, such as for dynamic control of an autonomous vehicle or a wind farm, or it may occur on slower time scales, such as post-flight updating of a digital twin for aircraft engine predictive maintenance or post-imaging updating of a digital twin and subsequent treatment planning for a cancer patient.

On the physical-to-virtual flowpath, digital twin tasks include sensor data fusion, model calibration, dynamic model updating, and estimation of parameters and states that are not directly observable. These calibration, updating, and estimation tasks are typically posed mathematically as data assimilation and inverse problems, which can take the form of parameter estimation (both static and dynamic), state estimation, regression, classification, and detection.

On the virtual-to-physical flowpath, the digital twin is used to drive changes in the physical counterpart itself or in the sensor and observing systems associated with the physical counterpart. This flowpath may be fully automated, where the digital twin interacts directly with the physical system. Examples of automated decision-making tasks include automated control, scheduling, recommendation, and sensor steering. In many cases, these tasks relate to automatic feedback control, which is already in widespread use across many engineering systems. Concrete examples of potential digital twin automated decision-making tasks are given in the illustrative examples in Box 2-1. The virtual-to-physical flowpath may also include a human in the digital twin feedback loop. A human may play the key decision-making role, in which case the digital twin provides decision support, or decision-making may be shared jointly between the digital twin and a human as a human–agent team. Human–digital twin interaction may also take the form of the human playing a crucial role in designing, managing, and operating elements of the digital twin, including selecting sensors and data sources, managing the models underlying the virtual representation, and implementing algorithms and analytics tools. User-centered design is central to extracting value from the digital twin.

Verification, Validation, and Uncertainty Quantification

VVUQ is essential for the responsible development, implementation, monitoring, and sustainability of digital twins. Since the precise definitions can differ among subject-matter areas, the committee adopts the definition of VVUQ used in the National Research Council report Assessing the Reliability of Complex Models (NRC 2012) for this report:

• Verification is “the process of determining whether a computer program (‘code’) correctly solves the equations of the mathematical model. This includes code verification (determining whether the code correctly implements the intended algorithms) and solution verification (determining the accuracy with which the algorithms solve the mathematical model’s equations for specified quantities of interest).”
• Validation is “the process of determining the degree to which a model is an accurate representation of the real world from the perspective of the intended uses of the model (taken from AIAA [Computational Fluid Dynamics Committee], 1998).”

• Uncertainty quantification is “the process of quantifying uncertainties associated with model calculations of true, physical quantities of interest, with the goals of accounting for all sources of uncertainty and quantifying the contributions of specific sources to the overall uncertainty.”

Each of the VVUQ tasks plays important roles for digital twins. There are, however, key differences. The challenges lie in the features that set digital twins apart from traditional modeling and simulation, with the most important difference being the bidirectional feedback loop between the virtual and the physical. Evolution of the physical counterpart in real-world use conditions, changes in data collection hardware and software, noisiness of data, addition and deletion of data sources, changes in the distribution of the data shared with the virtual twin, changes in the prediction and/or decision tasks posed to the digital twin, and evolution of the digital twin virtual models all have consequences for VVUQ. Significant challenges remain for VVUQ of stochastic and adaptive systems; due to their dynamic nature, digital twins inherit these challenges.

Traditionally, a computational model may be verified for sets of inputs at the code verification stage and for scenarios at the solution verification stage. While many of the elements are shared with VVUQ for computational models (NRC 2012), for digital twins, one anticipates, over time, upgrades to data collection technology (e.g., sensors). This may mean changes in the quality of data being collected, more and cheaper data capture hardware with potentially lower quality of information, different data sources, or changes in data structures. Additionally, the physical counterpart’s state will undergo continual evolution. With such changes comes the need to revisit some or all aspects of verification. Furthermore, as the physical twin evolves over its lifetime, it is possible to enter system states that are far from the solution scenarios that were envisioned at initial verification. Indeed, major changes made to the physical twin may require that the virtual representation be substantially redefined and re-implemented.

As with verification, validation is more complicated in the context of a digital twin. The output of a digital twin needs to include the confidence level in its prediction. Changes in the state of the physical counterpart, data collection and structures, and the computational models can each impact the validation assessment and may require continual validation. The bidirectional interplay between the physical and the virtual means the predictive model is periodically, or even continuously, updated. For continual VVUQ, automated VVUQ methods may yield operational efficiencies. These updates must be factored into digital twin validation processes.

Uncertainty quantification is essential to making informed decisions and to promoting the necessary transparency for a digital twin to build trust with decision support. Uncertainty quantification is also essential for fitness-for-purpose considerations. There are many potential sources of uncertainty in a digital twin. These include those arising from modeling uncertainties (Chapter 3),

measurement and other data uncertainties (Chapter 4), the processes of data assimilation and model calibration (Chapter 5), and decision-making (Chapter 6). Particularly unique to digital twins is inclusion of uncertainties due to integration of multiple modalities of data and models, and bidirectional and sometimes real-time interaction between the virtual representation, the physical counterpart, and the possible human-in-the-loop interactions. These interactions and integration may even lead to new instabilities that emerge due to the nonlinear coupling among different elements of the digital twin.

Given the interconnectedness of different systems and stakeholders across the digital twin ecosystem, it is imperative to outline the VVUQ pipeline and highlight potential sources of information breakdown and model collapse. It is important to recognize that VVUQ must play a role in all elements of the digital twin ecosystem. In the digital twin virtual representation, verification plays a key role in building trust that the mathematical models used for simulation of the physical counterpart have been sufficiently implemented. In cases that employ surrogate models, uncertainty quantification gives measures of the quality of prediction that the surrogate model provides. Field observations, for example, can be used to estimate uncertainties and parameters that govern the virtual representation (a type of inverse problem) as a step toward model validation, followed by the assessment of predictions. As information is passed from the physical counterpart to the virtual representation, new data can be used to update estimates and predictions with uncertainty quantification that can be used for decisions. These include challenges arising from model discrepancy, unresolved scales, surrogate modeling, and the need to issue predictions in extrapolatory regimes (Chapter 3).

When constructing digital twins, there are often many sources of data (e.g., data arising from sensors or simulations), and consequently, there can be many sources of uncertainty. Despite the abundance of data, there are nonetheless limitations to the ability to reduce uncertainty. Computational models may inherently contain unresolvable model form errors or discrepancies. Additionally, measurement errors in sensors are typically unavoidable. Whether adopting a data-centric or model-centric view, it is important to assess carefully which parts of the digital twin model can be informed by data and simulations and which cannot in order to prevent overfitting and to provide a full accounting of uncertainty.

The VVUQ contribution does not stop with the virtual representation. Monitoring the uncertainties associated with the physical counterpart and incorporating changes to, for example, sensors or data collection equipment are part of ensuring data quality passed to the virtual counterpart. Data quality improvements may be prioritized based on the relative impacts of parameter uncertainties on the resulting model uncertainties. Data quality challenges arise from measurement, undersampling, and other data uncertainties (Chapter 4). Data quality is especially pertinent when ML models are used. Research into methods for identifying and mitigating the impact of noisy or incomplete data is needed. VVUQ can also play a role in understanding the impact of mechanisms used to pass information

between the physical and virtual, and vice versa. These include challenges arising from parameter uncertainty and ill-posed or indeterminate inverse problems (Chapter 5). Additionally, the uncertainty introduced by the inclusion of the human-in-the-loop should be measured and quantified in some settings. The human-in-the-loop as part of the VVUQ pipeline can be a critical source of variability that also has to be taken into consideration (Chapter 6). This can be particularly important in making predictions where different decision makers are involved.

A digital twin without serious considerations of VVUQ is not trustworthy. However, a rigorous VVUQ approach across all elements of the digital twin may be difficult to achieve. Digital twins may represent systems-of-systems with multiscale, multiphysics, and multi-code components. VVUQ methods, and methods supporting digital twins broadly, will need to be adaptable and scalable as digital twins increase in complexity. Finally, the choice of performance metrics for VVUQ will depend on the use case. Such metrics might include average case prediction error (e.g., mean square prediction error), predictive variance, worse case prediction error, or risk-based assessments.

While this section has not provided an exhaustive list of VVUQ contributions to the digital twin ecosystem, it does serve to highlight that VVUQ plays a critical role in all aspects. Box 2-2 highlights the Department of Energy Predictive Science Academic Alliance Program as an exemplar model of interdisciplinary research that promotes VVUQ.

Conclusion 2-2: Digital twins require VVUQ to be a continual process that must adapt to changes in the physical counterpart, digital twin virtual models, data, and the prediction/decision task at hand. A gap exists between the class of problems that has been considered in traditional modeling and simulation settings and the VVUQ problems that will arise for digital twins.

The importance of a rigorous VVUQ process for a potentially powerful tool such as a digital twin cannot be overstated. Consider the growing concern over the dangers of artificial intelligence (AI),³ with warnings even extending to the “risk of human extinction” (Center for A.I. Safety 2023; Roose 2023). Generative AI models such as ChatGPT are being widely deployed, despite open questions about their reliability, robustness, and accuracy. There has long been a healthy skepticism about the use of predictive simulations in critical decision-making. Over time, use-driven research and development in VVUQ has provided a robust framework to foster confidence and establish boundaries for use of simulations that draw from new and ongoing computational science research (Hendrickson et al. 2020). As a result of continued advances in VVUQ, many of the ingredients

___________________

³ This report went through the review process prior to the October 30, 2023, release of the Biden-Harris administration Executive Order on the responsible development of artificial intelligence. While much of the discussion is relevant to this report, the committee did not have an opportunity to review and comment on the Executive Order as part of this study.

of AI methods—statistical modeling, surrogate modeling, inverse problems, data assimilation, optimal control—have long been used in engineering and scientific applications with acceptable levels of risk. One wonders: Is it the methods themselves that pose a risk to the human enterprise, or is it the way in which they are deployed without due attention to VVUQ and certification? When it comes to digital twins and their deployment in critical engineering and scientific applications, humanity cannot afford the cavalier attitude that pervades other applications of AI. It is critical that VVUQ be deeply embedded in the design, creation,

and deployment of digital twins—while recognizing that doing so will almost certainly slow progress.

Conclusion 2-3: Despite the growing use of artificial intelligence, machine learning, and empirical modeling in engineering and scientific applications, there is a lack of standards in reporting VVUQ as well as a lack of consideration of confidence in modeling outputs.

Conclusion 2-4: Methods for ensuring continual VVUQ and monitoring of digital twins are required to establish trustworthiness. It is critical that VVUQ be deeply embedded in the design, creation, and deployment of digital twins. In future digital twin research developments, VVUQ should play a core role and tight integration should be emphasized. Particular areas of research need include continual verification, continual validation, VVUQ in extrapolatory conditions, and scalable algorithms for complex multiscale, multiphysics, and multi-code digital twin software efforts.

Finding 2-2: The Department of Energy Predictive Science Academic Alliance Program has proven an exemplary model for promoting interdisciplinary research in computational science in U.S. research universities and has profoundly affected university cultures and curricula in computational science in the way that VVUQ is infused with scalable computing, programming paradigms on heterogeneous computer systems, and multiphysics and multi-code integration science.

Ethics, Privacy, and Data Governance

Protecting individual privacy requires proactive ethical consideration at every phase of development and within each element of the digital twin ecosystem. When data are collected, used, or traded, the protection of the individual’s identity is paramount. Despite the rampant collection of data in today’s information landscape, questions remain around preserving individual privacy. Current privacy-preserving methods, such as differential privacy or the use of synthetic data, are gaining traction but have limitations in many settings (e.g., reduced accuracy in data-scarce settings). Additionally, user data are frequently repurposed or sold. During the atmospheric and climate sciences digital twin workshop, for instance, speakers pointed out that the buying and selling of individual location data is a particularly significant challenge that deserves greater attention (NASEM 2023a).

Moreover, digital twins are enabled through the development and deployment of myriad complex algorithms. In both the biomedical workshop and atmospheric and climate sciences workshop on digital twins, speakers warned of the bias inherent in algorithms due to missing data as a result of historical and systemic biases (NASEM 2023a,b).

Collecting and using data in a way that is socially responsible, maintaining the privacy of individuals, and reducing bias in algorithms through inclusive and representative data gathering are all critical to the development of digital twins. However, these priorities are challenges for the research and professional communities at large and are not unique to digital twins. Below, the committee identifies some novel challenges that arise in the context of digital twins.

By virtue of the personalized nature of a digital twin (i.e., the digital twin’s specificity to a unique asset, human, or system), the virtual construct aggregates sensitive data, potentially identifiable or re-identifiable, and models that offer tailored insights about the physical counterpart. Speakers in the biomedical digital twin workshop remarked that a digital twin in a medical setting might include a patient’s entire health history and that a digital twin “will never be completely de-identifiable” (NASEM 2023b). As a repository of sensitive information, digital twins are vulnerable to data breaches, both accidental and malicious.

Speakers in both the biomedical workshop and the atmospheric and climate sciences workshop urged digital twin users and developers to enforce fitness for purpose and consider how the data are used. In a briefing to the committee, Dr. Lea Shanley repeated these concerns and stressed that the term “open data” does not mean unconditional use (Shanley 2023). During the atmospheric and climate sciences workshop, Dr. Michael Goodchild warned that “repurposing” data is a serious challenge that must be addressed (Goodchild 2023). Moreover, speakers highlighted the need for transparency surrounding individual data. As part of the final panel discussion in the biomedical workshop, Dr. Mangravite noted that once guidelines around data control are established, further work is needed to determine acceptable data access (Mangravite 2023).

The real-time data collection that may occur as part of some digital twins raises important questions around governance (NASEM 2023b). Dr. Shanley pointed out that using complex data sets that combine personal, public, and commercial data is fraught with legal and governance questions around ownership and responsibility (Shanley 2023). Understanding who is accountable for data accuracy is nontrivial and will require new legal frameworks.

Privacy, ownership, and responsibility for data accuracy in complex, heterogeneous digital twin environments are all areas with important open questions that require attention. The committee deemed governance to fall outside this study’s focus on foundational mathematical, statistical, and computational gaps. However, the committee would be remiss if it did not point out the dangers of scaling (or even developing) digital twins without clear and actionable standards for appropriate use and guidelines for identifying liability in the case of accidental or intentional misuse of a digital twin or its elements, as well as mechanisms for enforcing appropriate use.

Finding 2-3: Protecting privacy and determining data ownership and liability in complex, heterogeneous digital twin environments are unresolved challenges that pose critical barriers to the responsible development and scaling of digital twins.

Finally, making decisions based on information obtained from a digital twin raises additional ethical concerns. These challenges are discussed further in the context of automated and human-in-the-loop decision-making as part of Chapter 6.

Security

Characteristic of digital twins is the tight integration between the physical system and its virtual representation. This integration has several cybersecurity implications that must be considered, beyond what has historically been needed, in order to effectively safeguard and scale digital twins.

To maximize efficacy and utility of the digital twin, the physical counterpart must share as much of its data on a meaningful time scale as possible. The need to capture and transmit detailed and time-critical information exposes the physical system to considerably more risks. Examples include physical manipulation while feeding the digital twin fake data, misleading the operator of the physical counterpart, and intercepting data traffic to capture detailed data on the physical system.

As shown in Figure 2-1, feedback is integral to the digital twin paradigm. The close integration of physical and digital systems exposes an additional attack surface for the physical system. A malicious actor can inject an attack into the feedback loop (e.g., spoofing as the digital twin) and influence the physical system in a harmful manner.

An additional novel area of security consideration for digital twins arises from the vision of an ideal future where digital twins scale easily and effortlessly. Imagine the scenario where the digital twin is exposed to the broader community (either by design or inadvertently). Since the digital twin represents true physical traits and behaviors of its counterpart, malicious interactions with the digital twin could lead to security risks for the physical system. For example, consider the digital twin of an aircraft system; a malicious actor could manipulate the digital twin to observe vulnerable traits or behaviors of the physical system (e.g., because such traits or behaviors can be inferred from certain simulations, or changes in simulation parameters). These vulnerabilities may be unknown to the system operator. A malicious actor could also interrogate the digital twin to glean intellectual property data such as designs and system parameters. Therefore, scaling digital twins must take into consideration a balance of scalability and information sharing.

DIGITAL TWIN STATE OF THE ART AND DOMAIN-SPECIFIC CHALLENGES

During information-gathering sessions, the committee heard multiple examples of potential use cases for digital twins and some practical examples of digital twins being deployed. Use cases and practical examples arising in the domains of engineering, biomedical sciences, and atmospheric and climate sciences are summarized in the three Proceedings of a Workshop—in Brief (NASEM 2023a,b,c). Practical examples of digital twins for single assets and systems of assets are also given in a recent white paper from The Alan Turing Institute (Bennett et al. 2023). Digital twins can be seen as “innovation enablers” that are redefining engineering processes and multiplying capabilities to drive innovation across industries, businesses, and governments. This level of innovation is facilitated by a digital twin’s ability to integrate a product’s entire life cycle with performance data and to employ a continuous loop of optimization. Ultimately, digital twins could reduce risk, accelerate time from design to production, and improve decision-making as well as connect real-time data with virtual representations for remote monitoring, predictive capabilities, collaboration among stakeholders, and multiple training opportunities (Bochenek 2023).

While the exploration and use of digital twins is growing across domains, many state-of-the-art digital twins are largely the result of custom implementations that require considerable deployment resources and a high level of expertise (Niederer et al. 2021). Many of the exemplar use cases are limited to specific applications, using bespoke methods and technologies that are not widely applicable across other problem spaces. In part as a result of the bespoke nature of many digital twin implementations, the relative maturity of digital twins varies significantly across problem spaces. This section explores some current efforts under way in addition to domain-specific needs and opportunities within aerospace and defense applications; atmospheric, climate, and sustainability sciences; and biomedical applications.

Digital Twin Examples, Needs, and Opportunities for Aerospace and Defense Applications

There are many exciting and promising directions for digital twins in aerospace and defense applications. These directions are discussed in greater detail in Opportunities and Challenges for Digital Twins in Engineering: Proceedings of a Workshop—in Brief in Appendix E (NASEM 2023c); the following section outlines overarching themes from the workshop. The U.S. Air Force Research Laboratory Airframe Digital Twin program focuses on better maintaining the structural integrity of military aircraft. The initial goal of the program was to use digital twins to balance the need to avoid the unacceptable risk of catastrophic failure with the need to reduce the amount of downtime for maintenance

and prevent complicated and expensive maintenance. The use of data-informed simulations provides timely and actionable information to operators about what maintenance to perform and when. Operators can then plan for downtime, and maintainers can prepare to execute maintenance packages tailored for each physical twin and the corresponding asset (Kobryn 2023). The Department of Defense (DoD) could benefit from the broader use of digital twins in asset management, incorporating the processes and practices employed in the commercial aviation industry for maintenance analysis (Gahn 2023). Opportunities for digital twins include enhanced asset reliability, planned maintenance, reduced maintenance and inspection burden, and improved efficiency (Deshmukh 2023).

Significant gaps remain before the Airframe Digital Twin can be adopted by DoD. Connecting the simulations across length scales and physical phenomena is key, as is integrating probabilistic analysis. There is value in advancing optimal experimental design, active learning, optimal sensor placement, and dynamic sensor scheduling. These are significant areas of opportunity for development of digital twins across DoD applications. For example, by using simulations to determine which test conditions to run and where to place sensors, physical test programs could be reduced and digital twins better calibrated for operation (Kobryn 2023).

When building a representation of a fleet asset in a digital twin for maintenance and life-cycle predictions, it is important to capture the sources of manufacturing, operational, and environmental variation to understand how a particular component is operating in the field. This understanding enables the digital twin to have an appropriate fidelity to be useful in accurately predicting asset maintenance needs (Deshmukh 2023).

For DoD to move from digital twin “models to action,” it is important to consider the following enablers: uncertainty propagation, fast inference, model error quantification, identifiability, causality, optimization and control, surrogates and reduced-order models, and multifidelity information. Integrating data science and domain knowledge is critical to enable decision-making based on analytics to drive process change. Managing massive amounts of data and applying advanced analytics with a new level of intelligent decision-making will be needed to fully take advantage of digital twins in the future. There is also a need for further research in ontologies and harmonization among the digital twin user community; interoperability (from cells, to units, to systems, to systems-of-systems); causality, correlation, and uncertainty quantification; data–physics fusion; and strategies to change the testing and organizational culture (Deshmukh 2023; Duraisamy 2023; Grieves 2023).

Opportunities exist in the national security arena to test, design, and prototype processes and exercise virtual prototypes in military campaigns or with geopolitical analysis to improve mission readiness (Bochenek 2023).

Digital Twin Examples, Needs, and Opportunities for Atmospheric and Climate Sciences

Digital twins are being explored and implemented in a variety of contexts within the atmospheric, climate, and sustainability sciences. Specific use cases and opportunities are presented in Opportunities and Challenges for Digital Twins in Atmospheric and Climate Sciences: Proceedings of a Workshop—in Brief in Appendix C (NASEM 2023a). Key messages from the workshop panelists and speakers are summarized here. Destination Earth, or DestinE, is a collaborative European effort to model the planet and capture both natural and human activities. Plans for DestinE include interactive simulations of Earth systems, improved prediction capabilities, support for policy decisions, and mechanisms for members of the broader community to engage with its data (European Commission 2023). The models enabling DestinE are intended to be more realistic and of higher resolution, and the digital twin will incorporate both real and synthetic data (Modigliani 2023). The infrastructure required to support such robust and large-scale atmospheric, climate, and sustainability digital twins, however, necessitates increased observational abilities, computational capacity, mechanisms for large-scale data handling, and federated resource management. Such large-scale digital twins necessitate increased computational capacity, given that significant capacity is required to resolve multiple models of varying scale. Moreover, increasing computational abilities is not sufficient; computational capacity must also be used efficiently.

It is important to note that climate predictions do not necessarily require real-time updates, but some climate-related issues, such as wildfire response planning, might (Ghattas 2023). Three specific thrusts could help to advance the sort of climate modeling needed to realize digital twins: research on parametric sparsity and generalizing observational data, generation of training data and computation for highest possible resolution, and uncertainty quantification and calibration based on both observational and synthetic data (Schneider 2023). ML could be used to expedite the data assimilation process of such diverse data.

There are many sources of unpredictability that limit the applicability of digital twins to atmospheric prediction or climate change projection. The atmosphere, for example, exhibits nonlinear behavior on many time scales. As a chaotic fluid that is sensitively dependent on initial conditions, the predictability of the atmosphere at instantaneous states is inherently limited. Similarly, the physics of the water cycle introduce another source of unpredictability. The water phase changes are associated with exchanges of energy, and they introduce irreversible conditions as water changes phase from vapor to liquid or solid in the atmosphere and precipitates out to the Earth’s surface or the oceans.

The importance of—and challenges around—incorporating uncertainty into digital twins cannot be overstated. Approaches that rely on a Bayesian framework could help, as could utilizing reduced-order and surrogate models for tractability (Ghattas 2023) or utilizing fast sampling to better incorporate uncertainty (Balaji

2023). Giving users increased access to a digital twin’s supporting data may foster understanding of the digital twin’s uncertainty (McGovern 2023).

Establishing and maintaining confidence in and reliability of digital twins is critical for their use. One area for further development is tools that will assess the quality of a digital twin’s outputs, thus bolstering confidence in the system (NASEM 2023a). Predicting extreme events also poses challenges for widespread digital twin development and adoption. Because extreme events are, by definition, in the tail end of a distribution, methods for validating extreme events and long-term climate predictions are needed.

It is important to note that digital twins are often designed to meet the needs of many stakeholders, often beyond the scientific community. Using physics-based models in conjunction with data-driven models can help to incorporate social justice factors into community-centric metrics (Di Lorenzo 2023). It is necessary to include diverse thinking in a digital twin and to consider the obstacles current funding mechanisms pose toward the cross-disciplinary work that would foster such inclusion (Asch 2023).

Digital Twin Examples, Needs, and Opportunities for Biomedical Applications

Many researchers hold that digital twins are not yet in practical use for decision-making in the biomedical space, but extensive work to advance their development is ongoing. Many of these efforts are described in Opportunities and Challenges for Digital Twins in Biomedical Research: Proceedings of a Workshop—in Brief in Appendix D (NASEM 2023b). The European Union has funded various projects for digital twins in the biomedical space. The European Virtual Human Twin (EDITH)⁴ has the mission of creating a roadmap toward fully integrated multiscale and multiorgan whole-body digital twins. The goal of the project is to develop a cloud-based repository of digital twins for health care including data, models, algorithms, and good practices, providing a virtual collaboration environment. The team is also designing a simulation platform to support the transition toward an integrated twin. To prototype the platform, they have selected use cases in applications including cancer, cardiovascular disease, and osteoporosis. While questions for EDITH remain, including in the areas of technology (e.g., data, models, resource integration, infrastructure); users (e.g., access and workflows); ethics and regulations (e.g., privacy and policy); and sustainability (e.g., clinical uptake and business modeling) (Al-Lazikani et al. 2023), the work in this space is notable. DIGIPREDICT⁵ and the Swedish Digital

___________________

⁴ The website for the European Virtual Human Twin is https://www.edith-csa.eu, accessed June 30, 2023.

⁵ The website for DIGIPREDICT is https://www.digipredict.eu, accessed June 30, 2023.

Twin Consortium⁶ are two other examples of emerging European Union–funded projects working toward biomedical digital twins.

Technical challenges in modeling, computation, and data all pose current barriers to implementing digital twins for biomedical use. Because medical data are often sparse and collecting data can be invasive to patients, researchers need strategies to create working models despite missing data. A combination of data-driven and mechanistic models can be useful to this end (Glazier 2023; Kalpathy-Cramer 2023), but these approaches can remain limited due to the complexities and lack of understanding of the full biological processes even when sufficient data are available. In addition, data heterogeneity and the difficulty of integrating disparate multimodal data, collected across different time and size scales, also engender significant research questions. New techniques are necessary to harmonize, aggregate, and assimilate heterogenous data for biomedical digital twins (Koumoutsakos 2023; Sachs 2023). Furthermore, achieving interoperability and composability of models will be essential (Glazier 2023).

Accounting for uncertainty in biomedical digital twins as well as communicating and making appropriate decisions based on uncertainty will be vital to their practical application. As discussed more in Chapter 6, trust is paramount in the use of digital twins—and this is particularly critical for the use of these models in health care. Widespread adoption of digital twins will likely not be possible until patients, biologists, and clinicians trust them, which will first require education and transparency within the biomedical community (Enderling 2023; Miller 2023). Clear mechanisms for communicating uncertainty to digital twin users are a necessity. Though many challenges remain, opportunity also arises in that predictions from digital twins can open a line of communication between clinician and patient (Enderling 2023).

Ethical concerns are also important to consider throughout the process of developing digital twins for biomedical applications; these concerns cannot merely be an afterthought (NASEM 2023b). Bias inherent in data, models, and clinical processes needs to be evaluated and considered throughout the life cycle of a digital twin. Particularly considering the sensitive nature of medical data, it is important to prioritize privacy and security issues. Data-sharing mechanisms will also need to be developed, especially considering that some kinds of aggregate health data will never be entirely de-identifiable (Price 2023).

ADVANCING DIGITAL TWIN STATE OF THE ART REQUIRES AN INTEGRATED RESEARCH AGENDA

Despite the existence of examples of digital twins providing practical impact and value, the sentiment expressed across multiple committee information-gath-

___________________

⁶ The website for the Swedish Digital Twin Consortium is https://www.sdtc.se, accessed June 30, 2023.

ering sessions is that the publicity around digital twins and digital twin solutions currently outweighs the evidence base of success. For example, in a briefing to the committee, Mark Girolami, chief scientist of The Alan Turing Institute, stated that the “Digital Twin evidence base of success and added value is seriously lacking” (Girolami 2022).

Conclusion 2-5: Digital twins have been the subject of widespread interest and enthusiasm; it is challenging to separate what is true from what is merely aspirational, due to a lack of agreement across domains and sectors as well as misinformation. It is important to separate the aspirational from the actual to strengthen the credibility of the research in digital twins and to recognize that serious research questions remain in order to achieve the aspirational.

Conclusion 2-6: Realizing the potential of digital twins requires an integrated research agenda that advances each one of the key digital twin elements and, importantly, a holistic perspective of their interdependencies and interactions. This integrated research agenda includes foundational needs that span multiple domains as well as domain-specific needs.

Recommendation 1: Federal agencies should launch new crosscutting programs, such as those listed below, to advance mathematical, statistical, and computational foundations for digital twins. As these new digital twin–focused efforts are created and launched, federal agencies should identify opportunities for cross-agency interactions and facilitate cross-community collaborations where fruitful. An interagency working group may be helpful to ensure coordination.

• National Science Foundation (NSF). NSF should launch a new program focused on mathematical, statistical, and computational foundations for digital twins that cuts across multiple application domains of science and engineering.
  • The scale and scope of this program should be in line with other multidisciplinary NSF programs (e.g., the NSF Artificial Intelligence Institutes) to highlight the technical challenge being solved as well as the emphasis on theoretical foundations being grounded in practical use cases.
  • Ambitious new programs launched by NSF for digital twins should ensure that sufficient resources are allocated to the solicitation so that the technical advancements are evaluated using real-world use cases and testbeds.
  • NSF should encourage collaborations across industry and academia and develop mechanisms to ensure that small and medium-sized industrial and academic institutions can also compete and be successful leading such initiatives.

• • Ideally, this program should be administered and funded by multiple directorates at NSF, ensuring that from inception to sunset, real-world applications in multiple domains guide the theoretical components of the program.
• Department of Energy (DOE). DOE should draw on its unique computational facilities and large instruments coupled with the breadth of its mission as it considers new crosscutting programs in support of digital twin research and development. It is well positioned and experienced in large, interdisciplinary, multi-institutional mathematical, statistical, and computational programs. Moreover, it has demonstrated the ability to advance common foundational capabilities while also maintaining a focus on specific use-driven requirements (e.g., predictive high-fidelity models for high-consequence decision support). This collective ability should be reflected in a digital twin grand challenge research and development vision for DOE that goes beyond the current investments in large-scale simulation to advance and integrate the other digital twin elements, including the physical/virtual bidirectional interaction and high-consequence decision support. This vision, in turn, should guide DOE’s approach in establishing new crosscutting programs in mathematical, statistical, and computational foundations for digital twins.
• National Institutes of Health (NIH). NIH should invest in filling the gaps in digital twin technology in areas that are particularly critical to biomedical sciences and medical systems. These include bioethics, handling of measurement errors and temporal variations in clinical measurements, capture of adequate metadata to enable effective data harmonization, complexities of clinical decision-making with digital twin interactions, safety of closed-loop systems, privacy, and many others. This could be done via new cross-institute programs and expansion of current programs such as the Interagency Modeling and Analysis Group.
• Department of Defense (DoD). DoD’s Office of the Under Secretary of Defense for Research and Engineering should advance the application of digital twins as an integral part of the digital engineering performed to support system design, performance analysis, developmental and operational testing, operator and force training, and operational maintenance prediction. DoD should also consider using mechanisms such as the Multidisciplinary University Research Initiative and Defense Acquisition University to support research efforts to develop and mature the tools and techniques for the ap

• plication of digital twins as part of system digital engineering and model-based system engineering processes.
• Other federal agencies. Many federal agencies and organizations beyond those listed above can play important roles in the advancement of digital twin research. For example, the National Oceanic and Atmospheric Administration, the National Institute of Standards and Technology, and the National Aeronautics and Space Administration should be included in the discussion of digital twin research and development, drawing on their unique missions and extensive capabilities in the areas of data assimilation and real-time decision support.

As described earlier in this chapter, VVUQ is a key element of digital twins that necessitates collaborative and interdisciplinary investment.

Recommendation 2: Federal agencies should ensure that verification, validation, and uncertainty quantification (VVUQ) is an integral part of new digital twin programs. In crafting programs to advance the digital twin VVUQ research agenda, federal agencies should pay attention to the importance of (1) overarching complex multiscale, multiphysics problems as catalysts to promote interdisciplinary cooperation; (2) the availability and effective use of data and computational resources; (3) collaborations between academia and mission-driven government laboratories and agencies; and (4) opportunities to include digital twin VVUQ in educational programs. Federal agencies should consider the Department of Energy Predictive Science Academic Alliance Program as a possible model to emulate.

KEY GAPS, NEEDS, AND OPPORTUNITIES

In Table 2-1, the committee highlights key gaps, needs, and opportunities across the digital twin landscape. This is not meant to be an exhaustive list of all opportunities presented in the chapter. For the purposes of this report, prioritization of a gap is indicated by 1 or 2. While the committee believes all of the gaps listed are of high priority, gaps marked 1 may benefit from initial investment before moving on to gaps marked with a priority of 2.

TABLE 2-1 Key Gaps, Needs, and Opportunities Across the Digital Twin Landscape

REFERENCES