National Academies Press: OpenBook

Foundational Research Gaps and Future Directions for Digital Twins (2024)

Chapter: 6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities

« Previous: 5 Feedback Flow from Physical to Virtual: Foundational Research Needs and Opportunities
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

6

Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities

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 chapter identifies foundational research needs and opportunities associated with the use of digital twins for automated decision-making tasks such as control, optimization, and sensor steering. This chapter also discusses the roles of digital twins for providing decision support to a human decision-maker and for decision tasks that are shared jointly within a human–agent team. The chapter concludes with a discussion of the ethical and social implications of the use of digital twins in decision-making.

PREDICTION, CONTROL, STEERING, AND DECISION UNDER UNCERTAINTY

Just as there is a broad range of model types and data that may compose a digital twin depending on the particular domain and use case, there is an equally broad range of prediction and decision tasks that a digital twin may be called on to execute and/or support. This section focuses on tasks that manifest mathematically as control and optimization problems. Examples include automated control of an engineering system, optimized treatment regimens and treatment response assessments (e.g., diagnostic imaging or laboratory tests) recommended to a human medical decision-maker, optimized sensor locations deployed over an environmental area, automated dynamic sensor steering, and many more (see Box 6-1).

When it comes to these control and optimization tasks, a digital twin has unique features that challenge existing methods and expose foundational research

Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

gaps. Similar to the modeling challenges discussed in Chapter 3, the scale and complexity of a digital twin of a multiphysics, multiscale system (or system of systems) make control and optimization computationally challenging—even if decisions can be distilled down to a small number of quantities of interest, the decision variables (i.e., controller variables, optimization design variables) and system parameters affecting the control/optimization are likely to be of high dimension. Furthermore, the importance of digital twin verification, validation, and uncertainty quantification (VVUQ) is brought to the fore when it comes to decision-making tasks, yet the requisite end-to-end quantification of uncertainty places an even greater burden on the control and optimization methods. Lastly, the highly integrated nature of a digital twin leads to a need for tight and iterative coupling between data assimilation and optimal control—possibly in real time, on deployed computing platforms, and all with quantified uncertainty. Research gaps to address these challenges span many technical areas including operations research, reinforcement learning, optimal and stochastic control, dynamical systems, partial differential equation (PDE) constrained optimization, scalable algorithms, and statistics.

Rare Events and Risk Assessment in Support of Decision-Making

In many applications, digital twins will be called on to execute or support decisions that involve the characterization of low-probability events (e.g., failure in an engineering system, adverse outcomes in a medical intervention). In these cases, using digital twins to develop the quantifiable basis for decision-making requires a careful analysis of decision metrics, with particular attention to how one

Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

quantifies risk. A gap currently exists between state-of-the-art risk quantification methods and tools used for decision-making in practical science, engineering, and medicine contexts.

For example, risk metrics such as superquantiles are widely used for decision-making in the financial industry but have seen limited adoption in engineering (Royset 2023). The barriers go beyond just awareness and are in some cases systemic; for example, for some engineering systems, metrics such as probability of failure are encoded in certification standards. The challenges in assessing risk may be compounded in the context of digital twins developed to optimize one figure of merit but then adapted for a decision-making task in which performance metrics are different. Another challenge is that many risk measures lead to a non-differentiable objective. Chance constraints are often needed to impose probabilistic constraints on system behavior, resources, etc., and these can be non-differentiable as well. Monte Carlo gradient estimation procedures may be used to handle cases of non-differentiability. In general, non-differentiability complicates the use of gradient-based optimization methods, which in turn can limit scalability, and advanced uncertainty quantification methods relying on smoothness (e.g., stochastic Galerkin, stochastic collocation) may produce large integration errors. On the other hand, Monte Carlo sampling becomes extremely inefficient, especially when dealing with low-probability events.

Finding 6-1: There is a need for digital twins to support complex trade-offs of risk, performance, cost, and computation time in decision-making.

Sensor Steering, Optimal Experimental Design, and Active Learning

Within the realm of decision-making supported and executed by digital twins is the important class of problems that impact the data—specifically, the sensing and observing systems—of the physical counterpart. These problems may take the form of sensor placement, sensor steering, and sensor dynamic scheduling, which can be broadly characterized mathematically as optimal experimental design (OED) problems or in the data-driven literature as active learning. Just as for control problems, the needs for digital twins go beyond the capabilities of state-of-the-art methods.

Mathematically and statistically sophisticated formulations exist for OED, but there is a lack of approaches that scale to high-dimensional problems of the kinds anticipated for digital twins, while accounting for uncertainties and handling the rich modalities and complications of multiple data streams discussed in Chapter 4. Of particular relevance in the digital twin setting is the tight integration between sensing, inference, and decision-making, meaning that the OED problem cannot be considered in isolation. Rather, there is a need to integrate the OED problem with the data assimilation approaches of Chapter 5 and the control or decision-support task at hand (Ghattas 2022). That is, the sensors need to be

Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

steered in such a way as to maximize knowledge, not about the system parameters or state but about the factors feeding into the digital twin decision problem (e.g., the objective and constraints in an optimal control problem). The resulting integrated sense–assimilate–predict–control–steer cycle is challenging and cuts across several traditional areas of study but ultimately will lead to the most powerful instantiation of digital twins; therefore, scalable methods for goal-oriented sensor steering and OED over the entire sense–assimilate–predict–control–steer cycle merit further exploration.

Conclusion 6-1: There is value in digital twins that can optimally design and steer data collection, with the ultimate goal of supporting better decision-making.

Emphasizing the role of digital twins in data collection, it is crucial to recognize that they not only design and steer data gathering but also ensure the acquisition of high-fidelity, relevant, and actionable data. This capability enables more precise model training and fine-tuning, which translates to more reliable forecasts and simulations. Additionally, in the context of evolving environments and systems, digital twins play a pivotal role in adaptive data collection strategies, identifying areas that need more data and refining collection parameters dynamically. Ultimately, these capabilities, when harnessed correctly, can lead to more informed and timely decision-making, reducing risks and enhancing efficiency.

Digital Twin Demands for Real-Time Decision-Making

In several settings, control or optimization tasks may require real-time (or near-real-time) execution. The time scales that characterize real-time response may vary widely across applications, from fractions of seconds in an engineering automated control application to minutes or hours in support of a clinical decision. In many cases, achieving these actionable time scales will necessitate the use of surrogate models, as discussed in Chapter 3. These surrogate models must be predictive not just over state space but also over parameter space and decision variable space (Ghattas 2023). This places additional demands on both VVUQ and training data needs for the surrogates, compounding the challenges discussed in Chapter 3.

A mitigating factor can be to exploit the mathematical structure of the decision problem to obtain goal-oriented surrogates that are accurate with respect to the optimization/control objectives and constraints but need not reproduce the entire state space accurately. An additional challenge is that real-time digital twin computations may need to be conducted using edge computing under constraints on computational precision, power consumption, and communication. Machine learning (ML) models that can be executed rapidly are well suited to meet the computational requirements of real-time and/or in situ decision-making, but their black-box nature provides additional challenges for VVUQ and explain-

Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

ability. Additional work is needed to develop trusted ML and surrogate models that perform well under the computational and temporal conditions necessary for real-time decision-making with digital twins.

Finding 6-2: In many cases, trusted high-fidelity models will not meet the computational requirements to support digital twin decision-making.

Digital Twin Demands for Dynamic Adaptation in Decision-Making

A hallmark feature of digital twins is their ability to adapt to new conditions and data on the fly. In the context of decision-making, these new conditions may reflect, for instance, changes to the set of available states, the state transition probabilities themselves, or the environment. The nature of how decisions are made, particularly in automated control settings, may also necessitate dynamic adaptation. Reinforcement learning approaches address this setting, but currently there is a gap between theoretical performance guarantees for stylized settings and efficacious methods in practical domains. These challenges and gaps are exacerbated in the context of digital twins, where continual updates in response to new data require constant adaptation of decision-making methods. Safety-constrained reinforcement learning is beginning to address some of these issues for control problems in which it must be ensured that a system remains within a safe zone, particularly in the context of robotics and autonomous vehicles (Brunke et al. 2022; Isele et al. 2018).

Digital twins provide a useful mechanism for exploring the efficacy and safety of such methods. In addition, since the coupled data assimilation and optimal control problems are solved repeatedly over a moving time window, there is an opportunity to exploit dynamically adaptive optimization and control algorithms that can exploit sensitivity information to warm-start new solutions. Scalable methods to achieve dynamic adaptation in digital twin decision-making are necessary for exploiting the potential of digital twins. In the medical setting, beyond safety constraints are feasibility constraints; some treatment delivery recommendations may not be feasible (due to, for example, patient willingness or financial or system burden challenges) unless major changes are made to the entire system. For example, a recommendation of daily chemotherapy infusions could overwhelm the current system; however, home infusion capabilities may be a possible development in the future, and so feasibility constraints could also evolve over time.

Finding 6-3: Theory and methods are being developed for reinforcement learning and for dynamically adaptive optimization and control algorithms. There is an opportunity to connect these advances more strongly to the development of digital twin methodologies.

Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

Model-Centric and Data-Centric Views of Digital Twin Decision-Making

Underlying all these research gaps and opportunities to support digital twin decision-making is the synergistic interplay between models and data. As discussed in Chapter 3, a digital twin is distinguished from traditional modeling and simulation in the way that models and data work together to drive decision-making. The relative richness or scarcity of data together with the complexity and consequence of the decision space will significantly influence the appropriateness of different approaches (Ferrari 2023).

In data-rich scenarios, digital twins offer new opportunities to develop decision-making systems without explicit system models. However, to ensure that the resulting decisions are trusted, the digital twin must be able to not only predict how a system will respond to a new action or control but also assess the uncertainty associated with that prediction. Much of the literature on optimal decision-making focuses on incorporating uncertainty estimates; notable examples include Markov decision processes and bandit methods. However, these approaches typically reflect one of two extremes: (1) making minimal assumptions about the system and relying entirely on data to estimate uncertainties, or (2) placing strong assumptions on the model and relying on extensive calibration efforts to estimate model parameters a priori. Neither of these two approaches is well suited to incorporating physical models or simulators, and filling this gap is essential to decision-making with digital twins. Interpretability may also be a strong consideration, as first principles–based models may offer decision-makers an understanding of the model parameters and the causal relationships between inputs and outputs.

In data-poor scenarios, the models must necessarily play a greater role within the optimization/control algorithms. There are mature methods for deterministic optimization problems of this nature—for example, in the areas of model predictive control and PDE-constrained optimization. While advances have been made in the stochastic case, solving optimization problems under uncertainty at the scale and model sophistication anticipated for digital twins remains a challenge. A key ingredient for achieving scalability in model-constrained optimization is the availability of sensitivity information (i.e., gradients and possibly higher-order derivatives), often obtained using adjoint methods that scale well for high-dimensional problems. Adjoint methods are powerful, but their implementation is time intensive, requires specialized expertise, and is practically impossible for legacy codebases. Making sensitivity information more readily available would be an enabler for scalable decision-making with model-centric digital twins. This could be achieved by advancing automatic differentiation capabilities, with particular attention to approaches that will be successful for the multiphysics, multiscale, multi-code coupled models that will underlie many digital twins. Variational approaches that compute sensitivity information at the continuous (PDE) level are also emerging as promising tools.

Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

Finding 6-4: Models and data play a synergistic role in digital twin decision-making. The abundance or scarcity of data, complexity of the decision space, need to quantify uncertainty, and need for interpretability are all drivers to be considered in advancing theory and methods for digital twin decision-making.

HUMAN–DIGITAL TWIN INTERACTIONS

Human–computer interaction is the study of the design, evaluation, and implementation of computing systems for human use and the study of the major phenomena surrounding them (Sinha et al. 2010). Research and advances in interfaces and interactions of humans and computers continue to evolve considerably from the earliest Electronic Numerical Integrator and Calculator introduced in 1946 to modern graphical user interfaces (GUIs). However, the nature of human–digital twin interactions poses several unique challenges. The complex and dynamic nature of a digital twin introduces increased challenges around building trust and conveying evolving uncertainty, while also enabling understanding across all individuals who will interact with the digital twin. The contextual details required for digital twins can also introduce challenges in ethics, privacy, ownership, and governance of data around human contributions to and interactions with digital twins.

Use- and User-Centered Design

There is a range of respective roles that humans can play in interactions with digital twins, and the particular role and interaction of the human with a digital twin will influence the design of the digital twin. A key step is defining the intended use of the digital twin and the role of the human with the digital twin and clarifying the required design, development needs, and deployment requirements. The intended use of the digital twin along with the role and responsibilities of the human will help define the necessary data flows (which data at what time interval), range of acceptable uncertainties, and human–computer interaction requirements (e.g., how the data and uncertainties are presented).

In some settings, digital twins will interact with human operators continuously (as opposed to generating an output for subsequent human consumption and action). In these settings, input or feedback from human operators will dynamically alter the state of the digital twin. For example, one might consider a digital twin of a semi-autonomous vehicle that can solicit and incorporate human decisions. In such settings, care is needed to determine how to best solicit human feedback, accounting for human attention fatigue, human insights in new settings as well as human biases, and the most efficacious mechanisms for human feedback (e.g., pairwise or triplet comparisons as opposed to assigning raw numerical scores). Ongoing work in active learning and human–machine co-processing pro-

Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

vides many essential insights here. However, as with uncertainty quantification, existing methods do not provide clear mechanisms for incorporating physical or first-principle insights into decision-making processes, and closing this gap is essential for digital twins. To support human–digital twin interactions effectively, focused efforts must be made toward developing implementation science around digital twins, structuring user-centered design of digital twins, and enabling adaptations of human behavior.

Looking to the future, as emerging advances in the field of artificial intelligence (AI) allow for verbal and visual communication of concepts and processes, AI-mediated communications may be incorporated into digital twins to accelerate their creation, maintain their tight alignment with physical twins, and expand their capabilities. Moreover, a decision-making process could leverage a mixed team where AI “team members” manage large amounts of data and resources, provide classifications, and conduct analytical assessments. Human decision-makers could use this information to determine the course of action; in this way, AI components could assist in reducing the cognitive load on the human while enhancing the decision-making capabilities of the team.

Human Interaction with Digital Twin Data, Predictions, and Uncertainty

Effective visualization and communication of digital twin data, assumptions, and uncertainty are critical to ensure that the human user understands the content, context, and limitations that need to be considered in the resulting decisions. While opportunities for data visualization have expanded considerably over recent years, including the integration of GUIs and virtual reality capabilities, the understanding and visualization of the content in context, including the related uncertainties, remains difficult to capture; effective methods for communicating uncertainties necessitate further exploration.

Beyond the objective understanding of the uncertainties around the digital twin predictions, circumstantial and contextual factors including the magnitude of impact as well as the time urgency can influence human perception and decision-making amidst human–digital twin interactions. For instance, the prediction of daily temperature ranges for the week is likely to be received differently from the prediction and related uncertainty in the course of a hurricane or tornado, due to the magnitude of impact of the uncertainty and the immediate time-sensitive decisions that need to be made with this information. Similar parallels can be drawn for patients having their weight or blood pressure progress tracked over time amidst lifestyle or medication interventions, in which case some range of error is likely to be considered acceptable. In contrast, uncertainties in predictions of outcomes of interventions for a diagnosis of cancer would most likely be perceived and considered very differently by individuals due to the gravity of the situation and magnitude of impact of treatment decisions. While there is

Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

general recognition that the selected content and context (including uncertainties) around the presentation of information by the digital twin to the user will impact decision-making, there is limited research on the impact of the content, context, and mode of human–digital twin interaction on the resulting decisions. Uncertainty quantification advancements in recent decades provide methods to identify the sources of uncertainty, and in some settings an opportunity the reduce uncertainty (NRC 2012). Understanding uncertainty is not just a technical requirement but a foundational aspect of building trust and reliability among end users. When stakeholders are aware of the levels of uncertainty in the data or model predictions, they can make more nuanced and informed decisions. Additionally, the manner in which uncertainty is communicated can itself be a subject of research and innovation, involving the fields of user experience design, cognitive psychology, and even ethics. Techniques like visualization, confidence intervals, or interactive dashboards can be deployed to make the communication of uncertainty more effective and user-friendly.

Conclusion 6-2: Communicating uncertainty to end users is important for digital twin decision support.

Establishing Trust in Digital Twins

There are many aspects that add to the complexity of establishing trust for digital twins. As with most models, trust in a digital twin need not—and probably should not—be absolute. A digital twin cannot replace reality, but it might provide adequate insight to help a decision-maker. The end user needs to understand the parameter ranges in which the digital twin is reliable and trustworthy as well as what aspects of the digital twin outputs carry what level of trust. For example, a digital twin could be considered trustworthy in its representation of some physical behaviors but not of all of them. The interdisciplinary and interactive nature within the digital twin and across various stakeholders of the digital twin adds an additional layer of complexity to trust. Trauer et al. (2023) present three basic types of stakeholders in the context of a digital twin: the digital twin supplier, a user of the digital twin, and partners of the user. The authors describe how each of these stakeholders requires trust around different aspects related to the digital twin.

Similar to challenges of other artificial intelligence decision-support tools, interpretable methods are essential to establish trust, as humans generally need more than a black-box recommendation from a digital twin. However, an added complexity for digital twins is that as they change over time in response to new data and, in turn, experience changes in the recommended decisions, a way to communicate this evolution to the human user is needed—providing transparency of what has changed. For humans to believe and act on input from digital twins, the insights need to be presented with the acknowledged context of the physical

Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

elements including the constraints on the insights so that humans can understand and align the insights with the physical world. For example, a digital twin of electric grids should include the insights in the context of power flow mechanics in the electric grids.

As noted above, a key aspect to this transparency is the presentation of the insights with uncertainty quantification to the human in a manner that can be understood and considered in the decision-making process.

Finding 6-5: In addition to providing outputs that are interpretable, digital twins need to clearly communicate any updates and the corresponding changes to the VVUQ results to the user in order to engender trust.

Human Interactions with Digital Twins for Data Generation and Collection

The technology to collect data across the human–digital interface that could ultimately support human–digital twin interactions is growing rapidly. This includes using the digital twin software or interface to capture human interactions—for example, capturing the number of button clicks and mouse movements to facial expressions and gaze tracking. Data on human behavior such as biometrics being captured across various commercial devices that track step count, heart rate, and beyond can also inform digital twins for various applications to improve health care and wellness. With growing utilization of augmented reality and virtual reality, the collection of human interactions in the digital space will continue to increase and serve as a source of data for human–digital twin interactions. The data gathered within these interactions can also inform what and how future data capture is integrated into the digital twin (e.g., timing of assessments or measurements, introduction of new biosensors for humans interacting with digital twins).

As highlighted in other sections of this report, data acquisition and assimilation are major challenges for digital twins. Semantic interoperability challenges arise as a result of human–digital twin interactions. Moreover, data quality suffers from temporal inconsistencies; changes in data storage, acquisition, and assimilation practices; and the evolution of supporting technology. Despite various efforts across multiple organizations to establish standards, the adoption of standard terminologies and ontologies and the capture of contextual information have been slow even within domain areas. As we look to the multiscale, multidisciplinary interactions required for digital twins, the need to harmonize across domains magnifies the challenge. While the capture of enough contextual detail in the metadata is critical for ensuring appropriate inference and semantic interoperability, the inclusion of increasing details may pose risks to privacy.

Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

ETHICS AND SOCIAL IMPLICATIONS

Human–digital twin interactions raise considerations around ethics across various aspects including privacy, ethical bias propagation, and the influence of human–digital twin interactions on the evolution of human society. A digital twin of a human or a component (e.g., organ) of a human is inherently identifiable, and this poses questions around privacy and ownership as well as rights to access and the ethical responsibility of all who have access to this information. Individuals may be pressured or even coerced to provide the data being collected for the digital twin (Swartz 2018). For instance, in health care, the digital twin may be of a patient, but there are multiple humans-in-the-loop interacting with the digital twin including the patient and perhaps caregiver(s), providers that could encompass a multidisciplinary team, and other support staff who are generating data that feed into the digital twin. The governance around the data of all these interactions from all these humans remains unclear.

When considering human–digital twin interactions in health care, for instance, models may yield discriminatory results that may result from biases in the training data set (Obermeyer et al. 2019). Additionally, the developers may introduce biased views on illness and health into the digital twin that could influence the outputs. For example, a biased view grounded in a victim-blaming culture or an overly focused view on preconceived “health”-related factors that ignores other socioeconomic or environmental factors may clearly limit the ability to follow the recommendations to improve health (Marantz 1990). For example, patients who are not compliant with taking the recommended treatment may be viewed in a negative manner without considering potential financial, geographical, or other social limitations such as money to fill a prescription, time away from work to attend a therapy session, or even availability of fresh fruits and vegetables in their geographical proximity.

Conclusion 6-3: While the capture of enough contextual detail in the metadata is critical for ensuring appropriate inference and interoperability, the inclusion of increasing details may pose emerging privacy and security risks. This aggregation of potentially sensitive and personalized data and models is particularly challenging for digital twins. A digital twin of a human or component of a human is inherently identifiable, and this poses questions around privacy and ownership as well as rights to access.

Conclusion 6-4: Models may yield discriminatory results from biases of the training data sets or introduced biases from those developing the models. The human–digital twin interaction may result in increased or decreased bias in the decisions that are made.

Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

KEY GAPS, NEEDS, AND OPPORTUNITIES

In Table 6-1, the committee highlights key gaps, needs, and opportunities for enabling the feedback flow from the virtual representation to the physical counterpart of a digital twin. 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 6-1 Key Gaps, Needs, and Opportunities for Enabling the Feedback Flow from the Virtual Representation to the Physical Counterpart of a Digital Twin

Maturity Priority
Early and Preliminary Stages
Scalable methods are needed for goal-oriented sensor steering and optimal experimental design that encompass the full sense–assimilate–predict–control–steer cycle while accounting for uncertainty. 1
Trusted machine learning and surrogate models that meet the computational and temporal requirements for digital twin real-time decision-making are needed. 1
Scalable methods to achieve dynamic adaptation in digital twin decision-making are needed. 2
Theory and methods to achieve trusted decisions and quantified uncertainty for data-centric digital twins employing empirical and hybrid models are needed. 1
Methods and tools to make sensitivity information more readily available for model-centric digital twins, including automatic differentiation capabilities that will be successful for multiphysics, multiscale, multi-code digital twin virtual representations, are needed. 1
Research on and development of implementation science around digital twins, user-centered design of digital twins, and adaptations of human behavior that enable effective human–digital twin teaming are needed. Certain domains and sectors have had more success, such as in the Department of Defense. 1
Uncertainty visualization is key to ensuring that uncertainties are appropriately considered in the human–digital twin interaction and resulting decisions, but development of effective approaches for presenting uncertainty remains a gap. 2
While there is general recognition that the selected content and context (including uncertainties) around the presentation of information by the digital twin to the user will impact decision-making, there is limited research on the impact of the content, context, and mode of human–digital twin interaction on the resulting decisions. 1
Some Research Base Exists But Additional Investment Required
Scalable and efficient optimization and uncertainty quantification methods that handle non-differentiable functions that arise with many risk metrics are lacking. 2
Research Base Exists with Opportunities to Advance Digital Twins
Methods and processes to incorporate state-of-the-art risk metrics in practical science, engineering, and medicine digital twin decision-making contexts are needed. 2
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×

REFERENCES

Brunke, L., M. Greeff, A.W. Hall, Z. Yuan, S. Zhou, J. Panerati, and A.P. Schoellig. 2022. “Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning.” Annual Review of Control, Robotics, and Autonomous Systems 5:411–444.

Ferrari, A. 2023. “Building Robust Digital Twins.” Presentation to the Committee on Foundational Research Gaps and Future Directions for Digital Twins. April 24. Washington, DC.

Ghattas, O. 2022. “A Perspective on Foundational Research Gaps and Future Directions for Predictive Digital Twins.” Presentation to the Committee on Foundational Research Gaps and Future Directions for Digital Twins. November 15. Washington, DC.

Ghattas, O. 2023. Presentation to the Workshop on Digital Twins in Atmospheric, Climate, and Sustainability Science. February 1. Washington, DC.

Isele, D., A. Nakhaei, and K. Fujimura. 2018. “Safe Reinforcement Learning on Autonomous Vehicles.” Pp. 1–6 in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Madrid, Spain.

Marantz, P.R. 1990. “Blaming the Victim: The Negative Consequence of Preventive Medicine.” American Journal of Public Health 80(10):1186–1187.

NRC (National Research Council). 2012. Assessing the Reliability of Complex Models: Mathematical and Statistical Foundations of Verification, Validation, and Uncertainty Quantification. Washington, DC: The National Academies Press.

Obermeyer, Z., B. Powers, C. Vogeli, and S. Mullainathan. 2019. “Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations.” Science 366(6464):447–453.

Royset, J.O. 2023. “Risk-Adaptive Decision-making and Learning.” Presentation to the Committee on Foundational Research Gaps and Future Directions for Digital Twins. February 13. Washington, DC.

Sinha, G., R. Shahi, and M. Shankar. 2010. “Human Computer Interaction.” Pp. 1–4 in Proceedings of the 2010 3rd International Conference on Emerging Trends in Engineering and Technology (ICETET ‘10). IEEE Computer Society.

Swartz, A.K. 2018. “Smart Pills for Psychosis: The Tricky Ethical Challenges of Digital Medicine for Serious Mental Illness.” The American Journal of Bioethics 18(9):65–67.

Trauer, J., D.P. Mac, M. Mörtl, and M. Zimmermann. 2023. “A Digital Twin Business Modelling Approach.” Pp. 121–130 in Proceedings of the International Conference on Engineering Design (ICED23). Bordeaux: Cambridge University Press.

Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 85
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 86
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 87
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 88
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 89
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 90
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 91
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 92
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 93
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 94
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 95
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 96
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 97
Suggested Citation:"6 Feedback Flow from Virtual to Physical: Foundational Research Needs and Opportunities." National Academies of Sciences, Engineering, and Medicine. 2024. Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press. doi: 10.17226/26894.
×
Page 98
Next: 7 Toward Scalable and Sustainable Digital Twins »
Foundational Research Gaps and Future Directions for Digital Twins Get This Book
×
 Foundational Research Gaps and Future Directions for Digital Twins
Buy Paperback | $35.00
MyNAP members save 10% online.
Login or Register to save!
Download Free PDF

Across multiple domains of science, engineering, and medicine, excitement is growing about the potential of digital twins to transform scientific research, industrial practices, and many aspects of daily life. A digital twin couples computational models with a physical counterpart to create a system that is dynamically updated through bidirectional data flows as conditions change. Going beyond traditional simulation and modeling, digital twins could enable improved medical decision-making at the individual patient level, predictions of future weather and climate conditions over longer timescales, and safer, more efficient engineering processes. However, many challenges remain before these applications can be realized.

This report identifies the foundational research and resources needed to support the development of digital twin technologies. The report presents critical future research priorities and an interdisciplinary research agenda for the field, including how federal agencies and researchers across domains can best collaborate.

READ FREE ONLINE

  1. ×

    Welcome to OpenBook!

    You're looking at OpenBook, NAP.edu's online reading room since 1999. Based on feedback from you, our users, we've made some improvements that make it easier than ever to read thousands of publications on our website.

    Do you want to take a quick tour of the OpenBook's features?

    No Thanks Take a Tour »
  2. ×

    Show this book's table of contents, where you can jump to any chapter by name.

    « Back Next »
  3. ×

    ...or use these buttons to go back to the previous chapter or skip to the next one.

    « Back Next »
  4. ×

    Jump up to the previous page or down to the next one. Also, you can type in a page number and press Enter to go directly to that page in the book.

    « Back Next »
  5. ×

    Switch between the Original Pages, where you can read the report as it appeared in print, and Text Pages for the web version, where you can highlight and search the text.

    « Back Next »
  6. ×

    To search the entire text of this book, type in your search term here and press Enter.

    « Back Next »
  7. ×

    Share a link to this book page on your preferred social network or via email.

    « Back Next »
  8. ×

    View our suggested citation for this chapter.

    « Back Next »
  9. ×

    Ready to take your reading offline? Click here to buy this book in print or download it as a free PDF, if available.

    « Back Next »
Stay Connected!