“Interaction Twin in the middle”: a distributed digital twin architecture to model team interactions and dynamics for deep space missions

NASA’s Moon to Mars campaign emphasizes the need for crews and habitat systems to operate with increasing autonomy as communication delays with Earth grow beyond 5 minutes. The digital twin framework has emerged as a promising solution to monitor, diagnose, predict, and optimize space systems, but prior aerospace applications have largely centered on system autonomy rather than crew autonomy. As a result, current approaches under-represent the interaction dynamics needed by mission control to continuously evolve procedure and accomplish mission objectives. This work introduces an Interaction Digital Twin (IDT) framework that twins the interactions between humans and systems rather than focusing only on individual entities. Built on a distributed digital twin architecture with bidirectional information flow, the framework integrates three complementary types of twins: Digital Twins for habitat systems, Human Digital Twins (HDTs) for individual crew members, and Interaction Digital Twins that capture emergent phenomena such as team cohesion, trust calibration, coordination, and adaptive autonomy. Twinning the interactions moves aspects of command and control on-board, giving crew mission-control-like capabilities even during periods of communication delay. We apply the framework to an Artemis Phase II mission scenario, demonstrating how interaction-level twinning extends system-level modeling to support cognitive workload management, information sharing, and human–autonomy teaming. By elevating interactions to first-class, inference-capable elements within the digital twin architecture, this framework bridges the gap between technical system models and the human teaming constructs essential for self-sufficient deep space exploration.

The National Aeronautics and Space Administration (NASA) has proposed to establish a sustained human presence on and around the Moon, serving as a stepping stone for future deep space exploration, including missions to Mars (Harris et al., 2022). As part of this broader vision, NASA emphasizes the need for self-sufficient crew members and habitat systems to address the significant communication delay with Mars, which can reach up to 22 min one way, for a successful deep space mission (Drake, 2009b; a; Drake and Watts, 2014; Mindock et al., 2017; Stuster et al., 2019; Coan, 2020; Harris et al., 2022). Currently, astronauts aboard the International Space Station (ISS), which has less than a 10 s round trip delay, rely heavily on ground control for planning tasks, resolving anomalies, managing inventory, and coordinating missions, which align with higher level activities of monitoring, diagnosing, predicting, and optimizing systems (Parisi et al., 2023). However, even in this regime NASA notes that small communication delays (in the seconds range) can cause significant negative impact on performance, workload, and situational awareness (SA) (Fischer et al., 2013; Fischer and Mosier, 2014; Marquez et al., 2019). Furthermore, NASA recommends that for missions with communication delays exceeding 5 minutes, critical responsibilities must be transferred to the crew and systems onboard (Valinia et al., 2022; Parisi et al., 2023). Therefore, the future of spaceflight rests on the need for an onboard system that can monitor, diagnose, predict, and optimize for mission success.

To address this growing need for autonomy, the aerospace industry has increasingly adopted a digital twin framework as a promising solution for enhancing self-sufficiency in space systems (Allen, 2021). Originally developed for the manufacturing sector, Grieves defines a digital twin as “a set of virtual information constructs that fully describes a potential or actual physical manufactured product” (Grieves and Vickers, 2016). Since its introduction, this concept has been widely adopted and adapted across various domains, including advanced manufacturing (Davila-Gonzalez and Martin, 2024), human health and medicine (Tang et al., 2024), smart cities (Doctorarastoo et al., 2024), and aerospace (including human spaceflight) (Gratius et al., 2024; Sherman et al., 2024), with the goal of enabling greater autonomy. For example, the medical community uses digital twins to perform interventions aimed to answer ‘what-if’ questions, such as predicting how a patient would respond to a treatment without actually administering the drug (Tang et al., 2024).

However, prior applications of the digital twin framework in aerospace have primarily focused on enabling system autonomy rather than crew autonomy. The framework has demonstrated potential in deep space missions to reduce reliance on ground control, as evidenced by use cases across the literature that showcase its ability to monitor, diagnose, predict, and optimize systems (Gratius et al., 2024; Sherman et al., 2024). Existing work in general literature centers around: Digital twinning of a single human without also twinning the environment → human (Schmidt et al., 2025; Davila-Gonzalez and Martin, 2024)

Crowd or group twinning without modeling of individual level interactions → crew (Fan et al., 2021; Dembski et al., 2020)

Human to human interaction twinning with environment twinning, but in non-aerospace contexts, where individuals are not collaborating towards a unified objective, such as a mission → human-human in non-aerospace context (Doctorarastoo et al., 2024)

However, NASA has cited that the synergies (i) between the human and the system (human-system), (ii) between a team of humans and the system (crew-system), and (iii) among the human team (human-human) must also be considered to increase these capabilities for self-sufficiency (Coan, 2020; Mindock et al., 2017; Stuster et al., 2019; Harris et al., 2022; Drake, 2009b). While system autonomy contributes to crew autonomy, deep-space operations remain limited by the loss of time-critical cognitive and coordination functions traditionally performed by ground control. Ground control builds and evolves nominal operation procedures for crew based on their interactions (e.g., human-system, human-human) and the interactions’ results. Having the interactions twinned onboard and a Vehicle Systems Manager (VSM) that understands and synthesizes the interaction results would help augment nominal operations even with communication delays with mission control.

To address this gap in the literature, we propose a novel framework for human digital twinning, introducing specific elements that support deep space exploration by enabling greater crew autonomy. This framework is novel in the way that it addresses NASA’s design requirements for enhancing crew autonomy by twinning the interaction between crew members and their supporting systems, such as habitats and EVA suits, to support effective collaboration. Our framework is called an interaction digital twin (IDT) supporting three types of interactions: between an individual crew member and the system (human-system), among multiple crew members (human-human), and between a team of crew members and the system (crew-system)—all with increased cognitive workload management and collaboration as a central design objective (see Figure 1).

The remainder of the paper is organized as follows. Section 2 synthesizes the requirements identified by NASA for enabling self-sufficiency and increasing autonomy in human spaceflight. Section 3 introduces our proposed framework for interaction digital twins, enabling deep space missions. Section 4 provides a qualitative case study to assess the framework’s validity for deep space mission needs derived from literature.

Requirements specification

This section seeks to define the requirements that a digital twin framework must satisfy to support crew autonomy for deep space missions. Here, we survey self-sufficiency/awareness needs published by NASA to derive digital twin design requirements.

Human behavioral health and performance

In 2009, in preparation for the next stage of space exploration, NASA headquarters published a series of Design Reference Architectures (DRA) documents for human exploration of Mars (Drake, 2009a; Drake, 2009b; Drake and Watts, 2014). These DRAs are, to this day, some of the most recent and comprehensive documents to describe the capabilities (e.g., systems, infrastructure) and operations needed for a human deep space exploration mission. In addition to traditional engineering concerns, such as propulsion systems and in situ power generation, NASA emphasized the importance of information support provided by the habitat system, such as monitoring and diagnosing the habitat’s environmental health and delivering actionable insights to the crew. NASA noted that such capabilities are essential for reducing individual crew member stress, decreasing cognitive workload, and mitigating its downstream effects on team dynamics and overall mission performance.

Building on the Mars objective established in the DRA, NASA developed the Moon to Mars mission plan to define concrete steps for achieving beyond LEO travel. Here, we review research efforts published in response to NASA’s call to action.

The NASA Marshall Space Center expanded NASA’s reference architecture for the Moon to Mars mission plan by elaborating on Ground Rules and Assumptions in conjunction with functional allocations for the habitat design (Harris et al., 2022). While this document focuses more on the autonomy of the uncrewed habitat, it also details expectations of increased support from the habitat system for humans during the crewed portion of the flight (e.g., aiding crew awareness, supporting adaptability, facilitating purposeful social interactions). Also building upon the DRA, Johnson Space Center focused more on the human aspect of the Moon to Mars mission, releasing several publications that explicate the unique needs. Specifically, Mindock et al. (2017) listed 32 human system risk factors as critical areas for improvement to enable next-generation spaceflight. The risk factors that could be improved with crew-habitat teaming included crew autonomy, crew collaboration quality, human and vehicle automation integration, and information management support, which all subsequently decrease cognitive workload. Further, Stuster et al. (2019) outlined a comprehensive, ranked list of 1125 general tasks likely to be performed during the 12 phases of the first human expedition to Mars. Among the top 30 tasks included assessing displayed information from the system, performing health monitoring functions for habitat and crew, and adjusting system controls in response to displayed information. These tasks highlight the critical role of effective crew–habitat teaming, as many of the top-ranked tasks could be significantly enhanced through improved human–system collaboration. Finally, Coan (2020) highlighted key capabilities for exploration extravehicular activity (EVA) systems during the Artemis missions. The paper also emphasizes the need to enhance crew autonomy during Artemis missions by equipping EVA suits with an integrated in-suit informatics system, such as display and audio support for caution and warning alerts, mission progress updates, and crew activity coordination, to improve teamwork within the crew.

Together, these studies highlight the elements of the human-system interaction that are required for mission success (also found in Figure 2): Human-human interaction: improve crew collaboration quality to increase situation awareness and decrease cognitive workload

Crew-system interaction: support crew and system teaming

Human-system interaction: establish information management support from the system for individual crew members

System design requirements

The Findable, Accessible, Interoperable, and Reusable (FAIR) principles, originally introduced by Wilkinson et al. (2016), were established to guide a more sustainable and transparent data landscape. Recent work by Peters and Schindler (2024) has extended these 15 principles to the aerospace industry, emphasizing their role in enabling and enhancing digital twin systems. From their work we derive the following system design requirements for our framework: FAIR1: Enable shared vocabularies with unique identifiers.

FAIR2: Enable deep semantics by providing all information needed between any connection.

FAIR3: Reduce the complexity of data before transferring between system.

FAIR4: Enforce data provenance across the digital twin framework.

FAIR5: Data within the framework needs to be machine-interpretable.

FAIR6: Provide semantic, interconnected descriptions to enable knowledge-aware systems

Another guiding principle that we used to develop the proposed framework is NASA’s distributed hierarchical system architecture for future Earth-independent and autonomous deep space missions (Badger et al., 2019; Badger, 2018). This architecture is centered around a Vehicle Systems Manager (VSM) that provides system-wide autonomy and takes on responsibilities that are currently within NASA’s mission control center (MCC) on ground. The highest control authority still lies within the human operators (both crew and MCC). From the top-level control, inputs and commands are passed down the hierarchy to the lowest level. From the bottom up, devices, components, subsystems, systems, and modules provide health, status, and information to higher-level system managers (Badger et al., 2019; 2020). As the information from the lower-level is passed up, the complexity of data is reduced Sherman et al. (2024). To verify the proposed framework, it is essential to assess whether the interaction digital twin (IDT) framework can accommodate this architecture and represent the underlying physical system. For this purpose, we derived the following design requirements: Architecture1: Comply with fundamental principles of a distributed system architecture.

Architecture2: Enable the proper representation of the system architecture of the physical twin.

The last guiding principle is from understanding the role of mission control guiding off-nominal operation of missions. Historical spaceflight anomalies illustrate that one of the bottleneck under communication delay is the time-intensive cognitive and coordination functions performed by the ground control. During the Apollo 15 service propulsion system anomaly, over 75 verbal messages were passed from mission control to crew members, while in the Apollo 14 docking anomaly, more than 150 messages were exchanged Valinia et al. (2022). Similarly, during the December 2013 ISS cooling loop anomaly, more than 70 ground control personnel worked over several days to determine which subsystems should transition to contingency cooling, which loads could be safely powered down, and how recovery procedures should be sequenced to avoid cascading failures Valinia et al. (2022). These activities were not pre-scripted responses, but dynamic, distributed reasoning processes that evolved as crew actions produced new system observations.

Critically, the content of MCC-crew communication is as follows: different specialists reason about different subsystems, crew actions are executed to probe components, and the resulting system responses are jointly interpreted to revise the next set of actions. The procedure is not a static process, but an evolving plan shaped by interactions among people and systems. Under deep-space communication delays, this interaction loop cannot be sustained by the mission control, and should be moved onboard.

Literature review of existing human digital twin frameworks

As prefaced in the introduction, the aerospace industry has started to adopt the digital twin framework as a possible solution to increasing autonomy. In this section, we review the literature on digital twins across several important industries to better understand the capabilities available today. Here, the digital twin framework includes the digital twin, its corresponding physical twin, and the bidirectional information flow connecting the two. We adopt the Grieves/AIAA/AIA definition of a digital twin as a virtual representation that mirrors the structure, context, and behavior of the physical asset (Grieves and Vickers, 2016; AIAA Digital Engineering Integration Committee, 2020). The physical twin refers to the actual physical system or asset being twinned.

The digital twin framework relies on the bidirectional information flow: data from the physical twin is continuously sent to the digital twin, enabling it to reflect the current state of the system; in turn, the digital twin provides insights and recommendations that inform decisions and actions, which can then be implemented in the physical twin to realize value. The bidirectional information flow is often facilitated by data storage components, such as middleware, and a reasoning component, such as supervision, that processes information from both the physical and digital twin and decides what to do (e.g., which control input to insert into the physical twin) (Gratius et al., 2024; Sherman et al., 2024). Figure 3 below is an illustration of the definition we will be using throughout the paper.

In the remainder of this section, we review digital twin research across domains—including aerospace, healthcare, smart cities, and advanced manufacturing—with a focus on how these frameworks support autonomy through mirroring humans and their interaction with systems. These systems are generally representative of the space mission use case because they involve substantial physical and digital components as well as interaction between them.

Aerospace

Although digital twins for systems are widely discussed in aerospace, significantly less attention has been given to twinning humans (Tuegel et al., 2011; Glaessgen and Stargel, 2012). We highlight recent instances in the aerospace literature where human twinning has been attempted, albeit to a limited extent.

Gratius et al. (2024) defines a digital twin for a deep space Environmental Control and Life Support System (ECLSS) as comprising of four key components: The physical twin (e.g., the actual ECLSS hardware operating in the spacecraft), digital twin (e.g., a compilation of information and simulation models, supported by a software interface that organizes, processes, and retrieves data), middleware (e.g., a data storage and management layer that enables communication between the physical and digital components), and supervision (e.g., a decision-making component, consisting of algorithms or human operators). In Gratius et al. (2024), the supervision role was intended for either a human operator or a suite of algorithms. However, since Gratius et al. (2024) centers on an autonomous ECLSS, they focused on defining the suite of algorithms and its interaction with the digital twin to create control inputs.

Sherman et al. (2024) expands on this vision by introducing human-system teaming elements, including human cognitive state prediction, which involves estimating an individual crew member’s cognitive state to dynamically adjust the system’s level of autonomy, and crew operational state classification, which monitors an individual crew member’s workload capacity to ensure effective task allocation. While Sherman et al. (2024) focuses on the interaction between an individual crew member and the system, Schmidt et al. (2025) outlines the physiological information that should be monitored in a human body to enable medical risk assessment, prediction, personalization, clinical decision support, and recommendations.

The literature above demonstrates the potential for spaceflight system autonomy by adopting a digital twin framework, despite the relative de-emphasis of the human element. This contrasts with our next example, where humans are an intrinsic part of the digital twin framework.

Healthcare

The healthcare community primarily uses digital twins, often referred to as “Human Digital Twins”, to model humans for the purpose of experimenting with various forms of medical care. Due to the complexity of twinning a human, the community often recommends dividing human functions into at least two categories: physiological, which addresses the body’s physical processes, and psychological, which focuses on mental and emotional aspects (Chen et al., 2024a; Chen et al., 2024b; Chen et al., 2024c; Okegbile et al., 2023; Sahal et al., 2022). While the main focus being on providing care for a single human, interactions within healthcare DT’s is mostly aimed at the direct interaction between the different aspects of a single human’s DTs (e.g., cognitive DT with physiological DT) (Chen et al., 2024a; Chen et al., 2024b; Okegbile et al., 2023; Sahal et al., 2022; Vallée, 2023) There have been efforts (Vallée, 2023) to also model interactions between a single human DT and the environment or system that the patients are a part of (e.g., hospital setting); However, as mentioned before, most of the focus has been on accurately twinning a single human.

For the healthcare community, single human twinning enables key capabilities, such as real-time interventions (“what-if” simulations) on the digital twin without administering the potentially harmful changes onto the physical twin (patients). This allows for better diagnosis, telehealth, and clinical decision making through improvements in accuracy, time to diagnosis, and predictions (Erol et al., 2020; Meijer et al., 2023; Sun et al., 2023; Vallée, 2023). Additionally, the simulation capabilities are often identified as key enabling function for caregiver training and medication development/assessment (Alazab et al., 2023). For example, Erol et al. (2020) describes how through the integration of organ models built on the patients’ CT scans, real-time surgery support is enabled.

The literature above details works that focus mostly on twinning single humans. However, as identified in Section 2.1, we need a framework that can twin not only individual humans and systems, but the interaction between them. The next section steps towards modeling interaction as well as individual entities.

Smart cities

Smart cities (i.e., smart cities, cyber-physical-social systems, cyber-physical-social-infrastructure-systems) is a research area that aims to enhance quality of life and promote sustainability by leveraging digital technologies, communication systems, and data analytics (Gracias et al., 2023). In the smart city sector, digital twins, which use digital technologies and data analytics, often serve as collaborative platforms that enable stakeholders from traditionally siloed areas of the AEC industry, such as designers (e.g., city planners, architects), engineers (e.g., civil, mechanical), and builders (e.g., construction, project managers), to come together, assume the supervisory role, and make informed decisions about changes to the physical environment.

For example, Dembski et al. (2020) presents a digital twin of a German city in which human behavior is modeled as a group using movement potentials—where individuals tend to move from areas of high to low traffic. This digital twin integrates multiple data streams, such as 3D representations of the urban environment and traffic flow, enabling urban planners to visualize and supervise changes that improve the city’s livability. Yossef Ravid and Aharon-Gutman (2023) similarly positions humans in a supervisory role in their Social Urban Digital Twin for an Israeli city, which integrates data such as 3D spatial layouts, population demographics, and financial status. The model is used to understand how the elderly population interacts with social services, such as adult daycare facilities, and to identify ways to improve their experience.

Some works go beyond positioning humans in a supervisory role and instead embed them within the digital twin itself to model real-time decision making within the environment. In supervisory approaches, humans remain outside the digital model and are not represented in a way that captures their interactions and dynamic responses to environmental changes. By contrast, when humans are modeled within the digital twin, it becomes possible to observe how environmental modifications influence behavior and vice versa. For example, Ghansah and Lu (2023) proposes a framework for a “cognitive building”—a structure that continuously adapts to occupants’ needs to optimize performance and productivity—thus treating the human as an integral part of the digital model rather than an external decision-maker. Similarly, Ye and Wang (2022) proposes a digital twin to model cyber-physical-social interactions with the goal of minimizing unexpected human-system conflicts that may prevent systems from reaching their optimal state. Doctorarastoo et al. (2024) uses empirical data to develop a Bayesian Network model of group dynamics within a habitat, enabling the prediction of whether an interaction is likely to occur between two individuals. Fan et al. (2021) proposes a disaster city digital twin paradigm that integrates a meta-network of information, such as social, knowledge, task, and needs networks, alongside simulation models of various disaster scenarios. Within this framework, multiple actors interact with their respective systems to make decisions.

These examples incorporate humans into the digital twin framework primarily in supervisory roles, rather than modeling their real-time interactions with systems. The next section shifts focus to a domain that more directly emphasizes human–system interaction within the digital twin.

Advanced manufacturing

Many advances within the digital twin literature originate from developments in the advanced manufacturing domain. Modern manufacturing facilities are characterized by inherent machine complexity, automation, and the imperative to maintain safe and efficient workspaces for human operators. Consequently, many functional requirements in this domain overlap with the anticipated needs of human deep space exploration systems. In both contexts, there is a growing recognition that human performance is influenced by more than physiological well-being alone and increased automation/autonomous capabilities, necessitating a multimodal approach to twinning humans within complex systems. For example, Davila-Gonzalez and Martin (2024) proposed a HDT framework that integrates both physical and mental health metrics to ensure worker safety. This holistic approach to human twinning is further expanded by Lauer-Schmaltz et al. (2024), who identified the need for HDTs to encompass cognitive, mechanistic, physiological, and human-in-the-loop representations to accurately reflect the physical human twin within the system. However, these frameworks are primarily human-centric and do not incorporate non-human DTs within the architecture. In contrast, Bilberg and Malik (2019) developed a human-robot collaboration digital twin aimed at optimizing assembly tasks through high-fidelity simulation models of robotic systems. While effective in modeling the robotic system’s physical twin (e.g., trajectory planning), their architecture notably lacks integrated human models, limiting its applicability for fully representing human-system interaction dynamics.

Summary

To support deep space exploration, as outlined in Section 2.1, it is not only essential to develop individual digital twins for humans and systems (to monitor, diagnose, predict, and optimize) but also to model the interactions between these entities in order to reduce crew cognitive workload. Without such interaction digital twins, crew members must mentally account for these interactions between systems, substantially increasing cognitive workflow. Existing works across sectors such as aerospace, healthcare, smart cities, and manufacturing offer strong digital twin frameworks for individual entities, and some address specific interactions (e.g., human-human, human-system, or crew-system). However, as shown in Table 1, there is currently no unified framework capable of twinning all three types of interactions within a single architecture. To address this gap, we propose an Interaction Digital Twin (IDT) framework.

As future human exploration space missions operate with increasing communication delays and extended periods of Earth independence, crew autonomy must expand beyond procedural execution to encompass cognitive, coordination, and supervisory functions traditionally provided by ground control. Prior NASA work has emphasized that increasing onboard resilience, effectively enabling a small crew to achieve outcomes currently supported by dozens of ground-based system experts, requires the tight integration of new technologies that can assume these roles locally rather than relying on time-delayed ground interventions Valinia et al. (2023). Accordingly, to address these challenges, we propose an Interaction Digital Twin architecture that helps onboard essential ground control functions currently performed by flight directors, flight controllers, and backroom specialists, including mission cognizance, team communication and interaction support, procedure execution and adaptation, crew coordination, and qualitative performance monitoring Dempsey (2018).

As pointed out in Section 2, digital twins have been identified by the scientific community as a way to address many of these tasks and skills. However, while existing literature supports digital twinning of individual entities, it does not address the full spectrum of interactions expected and identified by NASA in long-duration missions, namely,: human-human, human-system, and crew-system. To fulfill this requirement, we present the Interaction Digital Twin framework, a digital twin architecture designed to support crew autonomy under deep space constraints by explicitly modeling interactions among mission-relevant entities.

Figure 4, the IDT, builds upon Figure 1, a single entity DT, while exploring the three interaction types to reduce cognitive workload, support coordination, and enable adaptive supervision without continuous ground involvement. It is important to note that the term ‘system’ may represent a collection of subsystems. For example, a single robot can be considered a subsystem and a coordinated fleet of such robots can constitute a larger robotic agent system. While subsystem-subsystem interactions are not the focus of this paper—having been addressed in Gratius et al. (2024) and Sherman et al. (2024), specifically for ECLSS—the interaction framework proposed here can be extended to model such relationships as well.

The subsequent subsections describe the individual components of the proposed IDT framework and how their design supports the crew autonomy requirements identified in Section 2. We align this discussion with the common elements of aerospace DT frameworks introduced by Gratius et al. (2024) and Sherman et al. (2024): Physical twin: The actual crew member and systems within the deep space habitat

Digital twin: A compilation of information and simulation models, supported by a software interface that organizes, processes, and retrieves data

Middleware: A data storage and management layer that enables communication between the physical and digital components

Supervision: A decision-making component, consisting of algorithms or human operators that interpret data from both twins and enact changes in the physical twin

Physical twin

While there is not yet a comprehensive system description for NASA’s future human deep space exploration missions, many of the key system elements can be identified from the literature. At this point in time, NASA is assuming a crew size of 4 (Harris et al., 2022). As discussed in detail in Section 2.1.1, these crews will operate under a completely new operational paradigm where onboard crew and systems will have unprecedented decision-making authority to overcome the challenges imposed by communication delays (Badger et al., 2020). To compensate for the loss of real-time support from Earth, these novel mission paradigms are anticipated to include highly integrated habitats that collect, process, share, and analyze data locally to enable autonomous mission operations (Badger, 2018; Harris et al., 2022; Drake, 2009a; b). Some of the commonly mentioned autonomous systems range from robotic agents (Harris et al., 2022; Fong, 2024) to completely virtual agents, such as vehicle management systems (Badger et al., 2019; Figueroa et al., 2020; Figueroa, 2023) or scheduling support systems (Marquez et al., 2017). Additionally, it is anticipated that astronauts during these missions will have several surface mobility vehicles available, like a pressurized rover that can be used as a temporary habitat (Drake, 2009a). Instead of building data silos for each individual mission element, NASA expects to accomplish the needed mission autonomy by integrating and sharing data across these systems (Badger et al., 2020; Harris et al., 2022). To build a proper DT system, all of the aforementioned mission elements (including crew members) are treated as physical twins that each have a corresponding digital representation (digital twin) within the DT system.

Digital twin

Next, we will describe the design and function of the DT itself. This element is different from the DT framework outlined in Section 2.2 above, as this section focuses on the digital twin component that is part of the complete (DT) framework.

The DT within our proposed framework contains the common 3 main DT elements 1) the software interface, 2) information models, and 3) simulation models (see Figure 5). The software interface is responsible for handling and routing queries placed from the supervision or other DT’s to the proper endpoints within itself. The software interface’s tasks can be summarized as task planning, model selection, task execution, and response delivery Gratius et al. (2024). Inside the DT there are two different model types: the information models, which are models that describe the physical twins’ attributes and are often used as inputs, configuration parameters, and/or contextual references for simulations; and the simulation models, which build the basis for enabling the DT’s ‘What-if’ simulation engine and replicate the physical twin’s dynamic behavior.

We propose the use of a distributed DT architecture. This DT architecture relies on the use of three different DT types to twin the whole habitat system (habitat systems and crewmember(s)). The three types of DT are: Digital Twin: A type of DT for non-human habitat systems and the environment

Human Digital Twins (HDT): A type of digital twin for each crew member

Interaction Digital Twin (IDT): A type of digital twin specifically for the interaction between habitat entities (i.e., human-human, human-to-system, crew-to-system) that captures the unique emergent team states and dynamics (expanded upon in Section 3.3) that are only existent during the interaction of multiple entities

Each of these types of DT’s are built on the same base fundamental DT architecture but do have different kinds of information models and simulation models. For example, as described by Schmidt et al. (2025) information models in HDT for astronauts should contain medical history of the specific crew members to enable proper HDT functionality. While a DT for a habitat system obviously does not contain a medical history but instead has a maintenance history (which in an abstract way can be interpreted as a medical history). The kinds of information and simulation models that (commonly) exist within DT’s and HDT’s have been extensively discussed in the existing literature, and have been summarized in Section 2.2.

Interaction digital twin (IDT)

The proposed IDT is responsible for twinning the constructs and phenomena that only occur when two or more system entities (human-human, human-system, crew-system) interact with one another. For example, during an EVA two astronauts work together on collecting geological samples from the Martian surface. During this activity, the two interacting astronauts build trust between each other through successful execution of the task. This team construct only exists and is measurable within the interaction, but its presence and development are essential for high team performance and task execution. Interaction constructs like this are captured within the IDT, thus, the information models of this DT type may contain team characteristics, team competency, and Ground Rules & Assumptions (GR&A) models. Complementing the information models are simulation models for team dynamics, which evolve around the emergent states that can change over time and/or through interactions, such as team cohesion, synchrony, and psychological safety. Further, it contains models focused on team performance, and coordination (e.g., adaptive autonomy control). The IDT applies to both human-human and human-system collaborations as concepts like ‘trust calibration’ exists in both team configurations. The combination of these information and simulation models enables the IDT to function as an information and reasoning library, providing context-aware assessments of team state, coordination, and performance that support supervisory decision-making by the crew and onboard systems. In this way the IDT contributes to an increase in crew autonomy by reducing reliance on continuous ground support. Table 2 outlines a comprehensive, though not exhaustive, set of models anticipated to form the foundational components of a DT system supporting future human deep space exploration missions derived from Gratius et al. (2024); Schmidt et al. (2025); Harris et al. (2024).

The proposed design for the IDT is functionally different to existing DT architectures proposed in the DT literature. Current architectures do describe collaborating twins in the context of distributed manufacturing by recognizing that complex systems will not be able to be represented by one simple digital twin but require multiple specialized twins. However, the collaboration is focused on information exchange between the DTs (Sahal et al. (2021); Sahal et al. (2022)) and lacks the twinning of interactions between systems. Other work described human-machine collaboration within a DT to twin the physical system but did not include the integration of independent DT’s to facilitate the collaboration and was demonstrated on a visual question answering system Wang et al. (2021).

IDT network architecture

Following the discussion of individual Digital Twin (DT) framework elements and their respective functions, this section presents the proposed DT network architecture designed to enable crew autonomy by accurately twinning human–human, human–system, and crew–system interactions. We first provide a high-level overview of these interaction architectures and highlight their distinctions, before examining each configuration in greater detail and identifying key architectural components in the context of deep space human exploration. While each interaction scenario is unique, the proposed framework adopts a modular design approach to ensure flexibility, scalability, compliance with NASA’s hierarchical system architecture, and adherence to FAIR (Findable, Accessible, Interoperable, Reusable) principles (Peters and Schindler, 2024).

The architecture for human–human interaction is illustrated in Figure 6. In this configuration, multiple human operators collaborate to form a human team. The human–human IDT utilizes the individual HDTs of the team members to accurately twin the physical system. This IDT represents a team size of 2 < n ≤ m, where n denotes the number of participating humans in a given interaction and m indicates the maximum number of possible teammates in the system.

In contrast, when a human operator interacts with a (habitat) system, a human–system IDT is instantiated, as shown in Figure 6. This configuration supports dyadic teaming scenarios and thus comprises only one HDT and one system DT.

Lastly, the crew–system interaction scenario depicted in Figure 6, which involves multiple humans interacting with each other and/or one or more systems. To maintain architectural modularity, this interaction is constructed by integrating the human–human IDT architecture with the human–system IDT, resulting in a scalable and modular crew–system IDT configuration that can be adapted to any interaction scenario encountered during Earth Independent Human Systems Operations. To reiterate, there is currently no unified framework capable of twinning all three types of interactions within a single architecture. To address this gap, we propose this IDT framework.

Human-human IDT

In this section, we describe the human–human IDT network architecture in more detail in the context of human deep-space exploration mission profiles. This configuration reflects the system architecture for a crew of four onboard a deep-space habitat. The architecture contains four individual Astronaut Digital Twins (ADTs), which are a type of Human Digital Twin (HDT). As described in Section 3.2, these ADTs provide personalized support by capturing each crew member’s unique physiological, psychological, cognitive, and behavioral attributes through both information and simulation models. Each astronaut’s personal data stream is used in the respective ADT to properly update and reflect the astronaut’s health and status.

In the architecture shown (Figure 7), all ADTs are connected to a Crew Digital Twin (CDT), a specialized form of IDT. This CDT represents the team’s emergent states, performance, and interactions that only exist when two or more astronauts meaningfully interact with each other. For example, team cohesion between astronaut 1 and astronaut 2 changes when both astronauts interact in some way. The collective measure of team cohesion for this dyad does not belong to either astronaut exclusively and, therefore, is not present in the individual ADTs; rather, it is captured and simulated within the CDT. In this instantiation, the CDT for anticipated human deep space exploration mission profiles operates as a 2 < n ≤ 4 IDT, where n represents the number of participating crew members.

The CDT is not simply a data-aggregation DT and thus differs in functionality from previous work, where a so-called collaboration twin connects two individual DTs to aggregate data from both into a single instance (Sahal et al., 2021; 2022). Instead, it is an inference-capable DT that may query other ADTs to access personalized data streams and contextual inferences for group-level assessments. For instance, to compute a crew synchrony metric during a dynamic task, the CDT may query individual ADTs for real-time heart rate data (data stream) or cognitive workload indicators (inference data). In this sense, the CDT functions analogously to a supervisory element, issuing requests to individual ADTs and using the responses to infer collective states and generate insights that inform system behavior or operational decisions. However, similar to other DTs in this proposed framework, the Crew DT does not have control authority over the physical twin but instead provides these inferences to the supervisory layer.

This layered and distributed architecture aligns with the hierarchical supervision principles discussed by Badger et al. (2019) and follows recommendations from Reiche et al. (2021) and Wang et al. (2022), Wang et al. (2024) for the comprehensive design of digital twin architectures. This also includes a designated DT software interface that manages the bidirectional communication between the CDT and each individual ADT. This interface governs the data flow and access rules, ensuring that only relevant and authorized information is passed from the ADT to the CDT while maintaining modular encapsulation.

This proposed networked approach to IDT and HDT design, where agent-level DTs feed into team-level DTs through managed interfaces, provides a scalable and modular path forward for modeling complex human–autonomy teaming phenomena.

Human-system IDT

Building upon the described CDT for human–human interactions, this section introduces the human–system IDT (Figure 8) in the context of astronauts and non-human teammates (e.g., robotic systems, autonomous agents, etc.) interacting during mission operations. For the purpose of describing the architecture in this section, we will use the example of a Robot Digital Twin (Robot DT), a type of DT that twins the physical onboard robotic system, as a representative non-human system. Based on this specific example, we therefore introduce a dedicated Human–Robot Interaction IDT (HRI IDT), which captures the teaming dynamics that emerge during human–robot interaction. The concepts of a Human-System IDT are, however, broadly applicable to any human-autonomy teaming scenario.

The core architecture of the HRI IDT remains consistent with the previously introduced CDT architecture, where each ADT represents the physical, behavioral, and performance attributes of its respective astronaut. However, in this configuration, a robot is also assigned a DT that reflects the state and performance of its physical twin. The Robot DT contains both information and simulation models (as described in Section 3.2) that mirror the physical and operational attributes of the physical system.

The IDT covering human–system interaction (the HRI IDT in this example) operates in the same manner as the CDT but is constrained to a dyadic relationship: one astronaut collaborating with one non-human system. The human–robot IDT contains dedicated HRI information and simulation models specialized to capture and simulate phenomena unique to mixed human–autonomy interactions, such as trust calibration, role negotiation, and dynamic task allocation.

All ADTs in the system have the potential to interface with the Robot DT via an IDT. However, each robot is represented by a single DT, and the interaction between an ADT and a specific robot-DT is structured to represent a single dyad—meaning that each teaming context is modeled discretely for each astronaut–robot pair. This architectural decision reflects operational realities in which astronauts interact with robotic teammates in distinct, task-specific roles. For example, astronaut 1 may collaborate with a robotic arm for external maintenance tasks, while astronaut 3 engages a mobile rover for geologic sampling—each scenario governed by a separate IDT.

The HRI IDT enables inference generation about the astronaut–robot dyad, such as assessing the dynamic evolution of trust or predicting latency in shared decision-making. Similar to the Crew DT, the HRI IDT does not have control authority but instead serves as an informational and predictive asset to the supervisory layer. Queries from the supervision, such as “Is this astronaut–robot team ready for unsupervised operation?”, can be routed through the software interfaces of the HRI IDT.

This modular approach to constructing the human–system IDT ensures that the unique characteristics of the physical twins are not conflated across different entities. For instance, the interaction of an astronaut operating the Canadarm onboard the ISS is fundamentally different from that of an astronaut working with a mobile robotics system like Astrobee, even though both may be classified as robotic systems.

Crew-system IDT

Figure 9 presents the final construct of the proposed IDT framework, combining the human–human interaction framework and the human–system interaction framework introduced above. This integrated architecture represents a fully modular and scalable network of DTs that supports human teams collaborating with non-human systems.

As in the previous architectures, each ADT twins one astronaut. These ADTs are linked to a CDT. Expanding beyond crew interactions, this final architecture includes a DT—in this example, a Robot DT—to represent the non-human teammate. Each human–system IDT, such as a Human–Robot IDT, operates under the same architectural principles as the CDT but is constrained to a one-to-one relationship, whereas the CDT can contain multiple ADTs.

This final architecture incorporates all principles and components described in the preceding sections, including bidirectional query routing through dedicated software interfaces, modular separation between information and simulation models, supervisory inference without direct control authority, and alignment with the layered design framework and NASA’s hierarchical distributed system architecture.

By integrating both human–human and human–system interaction models into a unified DT network, this architecture enables mission systems to reason about crew autonomy, system coordination, and team performance holistically. It provides a robust, extensible framework for dynamic decision support in future exploration-class missions.

The example shown in Figure 9 illustrates a single robotic teammate collaborating with the crew. However, the same principle applies when the crew (or an individual astronaut) interacts with multiple habitat systems during a task. For instance, if two astronauts require two or more robotic systems to accomplish a task, the IDT for those robotic systems interfaces directly with the Crew DT. For example, the Intra-Vehicular Robotics (IVR) IDT aggregates insights from the respective robotic system DTs to interact with the ‘Human Team to System Interaction IDT. This approach for creating an IVR DT is analogous to how the Crew DT is composed by integrating insights from all associated individual entity DT’s. Given the near-infinite combinations of future habitat systems and collaboration patterns, it is impractical to define every possible system DT in this paper. However, the proposed framework supports the formation of system-level digital twins that interact with each other, while providing flexibility during the system design phase.

IDT data architecture

The key differentiating characteristic of any DT is the utilization of data in real-time from the physical twin to create accurate inferences that serve the needs of mission control, crew, and the habitat. This is further emphasized in our proposed framework of the IDT, where data from different DT’s may be exchanged to allow cross-domain simulations. This section will briefly outline the data architecture of our proposed framework. To describe the proposed architecture, we utilize the layered architecture developed by Wang et al. (2024) and use the example of an astronaut-to-astronaut interaction (see Figure 10). The framework by Wang et al. (2024) consists of four layers: the agent layer, representing the physical twin; the data layer, responsible for data acquisition and management; the inference layer, which transforms raw data into actionable insights; and the implementation layer, which executes control or decision-making functions. Based on a difference in function allocation this approach has different DT elements associated at different layers. Nonetheless, the overall functionality is the same, which allows us to use the framework for our purposes.

Figure 10 highlights the flow of data within the IDT framework for a generic physical teaming task for two astronauts. In this example, the ECG data for both astronauts is fed into the data layer of the respective astronaut HDT’s. From there the respective HDT’s perform their regularly scheduled tasks. The task being collaborative in nature requires interaction between the two. For this generic example, the CDT (a type of IDT) requires the ECG data stream from the astronauts and crew psychological state to make inferences to the supervision. Thus, the processed ECG data stream from the data layer and crew cognitive state inference data from the inference layer of the astronaut DT is being queried and received into the data layer of the CDT. Once the CDT has all the data required for developing the interaction specific inference, the CDT delivers actionable information to the supervision.

In our proposed framework, individual DTs (i.e., DT and HDT) retain ownership of their respective entity-specific data. For instance, an ECG signal from astronaut 1 is managed and stored within the associated Astronaut 1 DT. When an interaction between astronaut 1 and another entity, such as astronaut 2, occurs, the IDT issues a query to retrieve the relevant datastream needed to generate appropriate inferences about the interaction. To support this process, the data must be forwarded from the Astronaut DT’s data layer to the interaction DT’s data layer. Additionally, configuring simulation models within the IDT may require inputs from the inference layer of the relevant ADT. This data flow and interaction sequence are further detailed in the sequence diagram shown in Figure 11.

Middleware and human datastreams

One of the key enabling and differentiating features of any DT is the integration of real-time datastreams of the physical twin. The isolated confined and extreme environment of spaceflight poses unique stressors to crew health and performance. NASA has historically taken a holistic approach to maintaining and monitoring human health in space which can be broadly categorized into physical, psychological, physiological, and social health (Pischulti et al., 2024). As there is currently no definitive set of expected datastreams for future deep space missions, this section outlines NASA’s existing approaches for collecting data on astronaut health and performance during missions in LEO. We then summarize commonly used sensing modalities identified in the broader research community that show promise for augmenting astronaut health and performance monitoring in deep space environments. Finally, we review key team dynamics and performance constructs from the literature. Understanding these constructs is particularly important for this work, as the proposed IDTs are anticipated to generate inferences about one or more of these constructs to enable real-time team interaction twinning based on existing human and habitat system datastreams.

Current human spaceflight operations

The current suite of sensors used to assess human health in LEO continues to evolve in response to the certification and deployment of new commercial spacecrafts, the development of novel sensor systems, changing mission profiles, and emerging scientific objectives. Historically, astronauts aboard the ISS have worn various wearable devices to monitor physiological signals such as heart rate, oxygen saturation, and body temperature. These signals are then used to derive common health related metrics like sleep patterns, stress levels, and physical activity (Gaskill, 2024a; b; Shahanaz, 2025). Despite these advancements, primary health assessments and care are still largely administered through Private Medical Conferences (PMCs) and Private Psychological Conferences (PPCs). These conferences allow crew members to confidentially discuss health-related concerns with the crew surgeon (Francisco, 2024). However, PMCs and PPCs are discrete events in which the data is collected and not a continuous datastream on the crewmembers physical and psychological health.

Beyond the PMCs, PPCs, and experimental wearable devices, NASA conducts a plethora of additional periodical health examinations including auditory, ocular, and cognitive assessments, body mass measurements, objective measures of sleep, and laboratory testing. The space agency additionally, places extra emphasis for crew health assessments during EVA by having real-time physiological measurements such as ECG and metabolic rate, as well as specialized medical examinations pre and post EVA (Francisco, 2024).

In addition to the aforementioned measurements, the ISS includes a vast array of system sensors that transmit real-time telemetry data to ground controllers. While the precise number of sensors and data streams is not publicly disclosed to the best of our knowledge, the ISS’ Caution and Warning System is known to comprise approximately 80 emergency alarms, 800 warnings, 2,300 cautions, and 6,100 advisories (Dempsey, 2018). While a small subset of these alerts are broadly applicable, the majority are specific to the ISS design and are expected to differ in future deep space habitat architectures.

Human sensing modalities for future human exploration missions

There are many potential modalities for monitoring human state during extended exploration missions. While many of these have been tested to extensively in ground-based exploration analog environments, relatively few have been tested in flight. We develop recommendations based on a combination of these study types, placing greater weight on the latter but using the former to inform the breadth of possibility. Two broad categories define the options for measurement: qualitative methods and quantitative methods. Beyond the standard subjective self-assessments (e.g., surveys and interviews), qualitative methods also include behavioral measurements that attempt to extract more objective data from participants through observation of behaviors. This includes motion tracking—usually using room-mounted motion capture systems or wearable inertial measurement units (IMUs)—to characterize physical behavior during interaction, structured observations by an experimenter to note elements of behavior like interactions with others, and vocal monitoring to capture both lexical and prosodic changes during interaction. These behavioral measures can infringe on crewmember privacy, so their use has been relatively limited in flight.

Quantitative methods generally rely on the collection of signals produced by the human body during activity, falling under two categories: psychophysiological monitoring and neurophysiological monitoring. Psychophysiological monitoring focuses on body systems connected to the autonomic nervous system that are controlled subconsciously (i.e., the “fight or flight” response), while neurophysiological measures focus on signals produced by the brain that are indicative of different cognitive functions. Psychophysiological sensing is generally more robust in noisy conditions than neurophysiological sensing, but neurophysiological sensing provides more precision in characterizing complex constructs like cognitive workload, trust, and situation awareness. Common measures are derived from the cardiovascular system (heart rate; HR, heart rate variability; HRV, and blood pressure; BP), respiratory system (respiration rate, oxygen saturation), integumentary system (electrodermal activity—i.e., sweat gland activation, skin temperature), musculoskeletal system (electromyography—muscle activation), the ocular system (pupillometry, gaze tracking), and the brain (functional near infrared spectroscopy; fNIRS, electroencephalography; EEG). Each modality has demonstrated utility in tracking changes in human state outside of the space exploration environment. HR, HRV, and BP have been shown to change with stress, cognitive load, and emotional state and are relatively non-invasive and unobtrusive measures Rainville et al. (2006), Arquilla et al. (2022). Similarly, respiration rate and respiration rate variability can be good indicators of arousal and workload Veltman and Gaillard (1998). Respiration measures have been used successfully in several studies relating to measuring stress in human robot interactions Novak et al. (2015). Sweat gland activation and skin temperature have also been used as indicators of arousal in studies of human-robot and human-system interaction, and skin temperature can be measured using thermal cameras, reducing the burden of on-body sensing, Kulic and Croft (2005), Sorostinean et al. (2015). EMG sensing is useful for tracking physical activity and physical interaction, and the ocular and brain measurements are most useful for tracking attention and cognitive workload during human-system interaction.

The challenge with multi-modal sensing for human state estimation within a digital twin stems from the fact that each signal is produced by a different part of the body with variability in response latency and recovery times. The same sampling frequency across all signals will lead to missed information in some signals and overreliance on others. Similarly, selecting appropriate epochs for feature extraction depends on the unique combination of sensing modalities and some knowledge of the types of stressors that will be encountered by the human. These challenges necessitate tuning of the approach to human monitoring based on the use case and potentially the specific crewmembers involved to develop accurate, reliable, and robust data for an effective digital twin.

Team dynamics and performance

NASA has consistently emphasized the importance of a well-functioning, high-performing crew to ensure mission success and prevent behavioral health decrements during human spaceflight operations (NASA Human Research Program, 2025). This extended time spent together not only allows the crew to form strong interpersonal relationships with each other, but also enables them to learn behavioral patterns, and collaborative procedure execution. Over the course of a mission (and repeated interactions) team dynamics shift and may alter overall team performance (Landon et al., 2018).

Recent efforts by the NASA Behavioral Health and Performance Lab have been focused on standardizing “exploration measures” of team dynamics and performance, but there is not a current, published standard for measurement of these constructs across environments and in particular for Earth-independent human systems operations (Harris et al., 2024; NASA Human Research Program, 2025). Nevertheless, the team dynamics and human-autonomy teaming literature offers several well-validated constructs. One such construct is team cohesion, defined as “a shared attraction, bonding, or sense of pride among team members that is driven by social- or task-based elements” (Grossman et al., 2022), which has been shown to positively correlate with team performance (Driskell et al., 2024; Grossman et al., 2022; Guastello et al., 2022). Another extensively studied construct is trust, particularly in the context of human-autonomy teaming, where the operator’s trust in an autonomous agent influences task performance, decision-making, and system interaction patterns. Both over-reliance and under-trust have been linked to degraded performance outcomes, highlighting the need for constant attention (Duan et al., 2024). Closely related, yet distinct, is the concept of psychological safety, defined as the “shared belief held by members of a team that the team is safe for interpersonal risk taking” (Edmondson, 1999). Unlike trust, which is often viewed as dyadic and directed, psychological safety reflects a shared team-level norm that fosters open communication and constructive risk-taking and is highly correlated with high-performing teams. These constructs (cohesion, trust, and psychological safety) collectively support team performance by strengthening interpersonal, and often parasocial, connections among team members. A complimentary set of team dynamic concepts focuses on team member alignment, coordination, and adaptation. For example, synchrony, shared situation awareness, and team cognition were all found to be associated with team performance and highlighting the importance of team member alignment (Cooke et al., 2013; Gorman et al., 2025).

The vast majority of well-validated team dynamics measures rely on external subject matter expert assessments (Edmondson, 1999), subjective survey (Edmondson, 1999; McComb et al., 2017) measures, and task-specific performance metrics (McComb et al., 2017). This makes real-time quantification of the current state of a team expensive, and especially for deep space missions with significant communication delays, often unfeasible. However, recent advances in the Internet of Things (IoT), human activity recognition through AI, and the availability of low-cost wearable sensors (and well-established individual state measurements derived from physiological signals) have enabled a growing body of research aimed at developing unobtrusive, real-time indicators of team dynamics across diverse sensing modalities (Landon et al., 2018; Gorman et al., 2020; 2025).

Supervision

The supervision element within the DT framework represents an integral part of crew autonomy as this is the area in which the decision-making process occurs and control inputs are passed into the physical twin. The supervision element serves as the primary interface through which system operators and autonomous decision-making processes interact with the digital twin. Figure 12 below shows how the distributed hierarchical framework as proposed by Badger et al. (2019) is represented within the supervision. As described in Section 3.1 the elements with higher-levels of authority pass control and command inputs down the chain of command; lower level elements within this framework provide health, status, and information to higher level system elements.

Human operators (MCC and crew) have task-specific user interfaces available to them. However, these interfaces do vary from each other to account for the difference in knowledge and expertise. For example, ground controllers responsible for thermal control are highly trained and have years of experience in operating this system, thus, their supervision interface has different requirements compared to crew when they try to answer what-if scenarios related to the thermal system. Similarly, habitat systems like the VSM do not require a user interface but rather interact with the DT through an API gateway that handles routing of the queries.

Due to the nature of the distributed hierarchical framework queries may not always be placed by the entity that ultimately makes the final decision. By requiring the health, status, and information to flow upward in the architecture, a low-level entity like the subsystems manager will place a query to their corresponding DT to provide answers about the anticipated behavior of the system during a dynamic event. The response or interpretation of these results are then passed to the Systems Manager (SM), which can then either pass this information upwards (e.g., to the MS) or use this information to place a query themselves.

In the following section, we conduct a partial verification of the proposed IDT framework by first examining its alignment with the FAIR principles, followed by an assessment of its compatibility with NASA’s distributed hierarchical system architecture. As a formal validation is not yet feasible, we instead discuss how the framework addresses the anticipated needs of future exploration crews. Finally, we outline identified gaps and limitations of the current work.

Framework verification

Alignment with FAIR principles

The FAIR (Findable, Accessible, Interoperable, and Reusable) principles, originally introduced by Wilkinson et al. (2016), were established to guide a more sustainable and transparent data landscape. Recent work by Peters and Schindler (2024) has extended these principles to aerospace settings, emphasizing their role in enabling and enhancing digital twin systems. In particular, their analysis highlights practical approaches for integrating FAIR principles within the industrial context. Building on these insights, we will discuss how our proposed IDT framework aligns these FAIR-oriented requirements.

Shared semantics (FAIR 1 and FAIR2)

A key aspect within FAIR the principles is to have shared vocabularies. While the implementation of these unified vocabularies is obviously highly dependent on close coordination within the engineering team to prevent the duplication of parameters, our implementation within the IDT, which makes the data entity-centric (e.g., ECG datastream of astronaut 1, is only available through the astronaut 1 DT) and requires a query to be accessed by other DT’s within the system, enforces the use of a shared vocabulary to a certain extent. Further, Peters and Schindler (2024) identified that the system should enable deep semantics to enable multiple layers of different granularity. Our framework enables this by modeling each individual entity in the system with its own DT and subsequently allowing DT’s to query each others simulation and information models. This enables deep granular information about any given system within the habitat.

Metadata (FAIR3)

Another critical implementation strategy involves the use of metadata to minimize redundant transfers of large datasets. While our framework does not mandate a specific data format or processing steps, the IDT architecture facilitates direct querying of actionable information from the inference layers of other DT. This design supports efficient data exchange without duplicating storage or transmission efforts. Moreover, by enabling inference-level access (for actionable data) rather than unrestricted data stream access, the framework ensures that control over sensitive information remains with the individual DT, thereby maintaining data governance. Access to raw data within the framework requires explicit authorization from the respective data owner.

Provenance (FAIR4)

Understanding the origin and provenance of data represents another key FAIR-enabling implementation approach. The proposed framework incorporates digital twins for all major physical entities within the system, each maintaining a complete record of the component’s state and history across its life cycle. Beyond physical entities, the framework extends this concept to include interaction-level (teaming) occurrences, ensuring that activities without a distinct physical representation, such as human collaboration with onboard robotic systems, are also captured and documented. This comprehensive approach preserves the contextual history of both physical and non-physical interactions, thereby enhancing traceability, accountability, and overall data re-usability within the digital twin network.

Open formats (FAIR5)

To enhance interoperability, the adoption of open formats is recommended to ensure that actionable information remains machine-interpretable. As noted previously, our framework does not prescribe specific data formats or processing standards. Instead, it requires that all data exchanges between digital twins occur through standardized queries via their respective software interfaces. This design choice underscores the importance of engineering teams adopting common and open data formats, thereby facilitating seamless machine-to-machine communication and interoperability across the digital twin network.

Knowledge-aware systems (FAIR6)

The IDT framework realizes the semantic interoperability envisioned by FAIR by embedding structured, machine-actionable descriptions within the information models of the DTs at every system level. Through standardized query interfaces, each DT can interpret and access semantically linked data across other DTs. Furthermore, the framework introduces IDTs that capture and link the semantics of system interactions, maintaining a historical record within their information models that forms the foundation for knowledge graphs focused on human–autonomy teaming. This approach moves the IDT toward a semantically integrated network of interoperable DTs, substantially reducing reliance on human subject-matter expertise.

Alignment with distributed hierarchical system architecture

In Section 3.1, we described the distributed hierarchical system architecture proposed by Badger et al. (2019), which is envisioned to support NASA’s future deep space exploration missions through the distribution of habitat functions and the integration of onboard autonomy capabilities. To verify the proposed framework, it is essential to assess whether the IDT framework can be effectively integrated within this architecture and whether it can adequately represent the underlying physical system.

Compliance with fundamental principles of a distributed architecture (Architecture1)

As the name suggests, a defining characteristic of a distributed architecture is the absence of monolithic structures. In the proposed framework, DTs map directly to their respective physical counterparts, ensuring modularity and clear system boundaries. Interactions between entities are captured by separate DTs that specialize in specific instances of interaction. For example, distinct IDTs are employed to represent human–human and human–system interactions, each further tailored to the particular system or subsystem being twinned. A key advantage identified in distributed architectures is the reduction of data complexity as information propagates through the system (Badger et al., 2019; 2020; Sherman et al., 2024). The proposed framework addresses this principle by allowing DTs to query other DTs within the network rather than directly exchanging full data streams. Information retrieved from a DT’s inference layer is inherently less complex than the raw signals originally obtained from the physical twin, thereby minimizing data transfer requirements and enhancing overall system efficiency.

Is it hierarchical (Architecture2)

The proposed framework is not inherently hierarchical in its structure. However, because it mirrors the organization of the physical system, it naturally adopts the hierarchical configuration of the system it twins. The semantic representation of this hierarchy is preserved within the information models of each individual DT, ensuring that relationships and dependencies between subsystems are faithfully maintained. Accordingly, the framework complies with the requirement to represent and integrate with hierarchical system architectures. It is important to note that many hierarchical aspects discussed by Badger et al. (2020), particularly those concerning locus of control, primarily pertain to the design of the supervision element and area beyond the scope of this work. Nevertheless, Section 3.4 discussed relevant aspects of the supervision element to our proposed framework in Section 3.4.

Preliminary framework validation

In Section 2.1, we identified three mission-enabling requirements for future human exploration missions: (i) improving crew collaboration quality, (ii) supporting crew–habitat teaming, and (iii) establishing information management mechanisms for monitoring system health. While the proposed IDT framework does not introduce new data-driven models, it serves as an enabling digital twin architecture that provides human behavioral health and performance focused models with needed data access and contextual information. Further, this capability supports scenario-based prediction of behavioral health and performance outcomes that emerge from human–system interactions. Most importantly, the framework establishes a platform capable of accurately twinning interaction scenarios between entities; scenarios that would otherwise be missing or violate design constraints (see Section 4.1). As summarized in Table 1, existing DT and HDT approaches satisfy only subsets of the identified requirements and generally lack the ability to accurately twin the complex system interactions found in human deep space mission architectures, limiting their ability to support crew autonomy under communication delay.

The IDT addresses this gap by explicitly modeling the interaction structure that underpins ground-supported anomaly response, rather than focusing solely on the state evolution of individual systems or crew members. By representing (i) human–system interactions (who is responsible for which subsystem and how actions affect system state), (ii) human–human interactions (how responsibilities, information, and decisions are coordinated across crew roles), and (iii) crew–system interactions (how collective actions propagate across coupled subsystems), the IDT enables the VSM to conduct onboard synthesis of mission state, dynamic responsibility allocation, and adaptive sequencing of actions without continuous ground input.

For example, during a cooling loop anomaly, the VSM, with the IDT, can reason over competing responsibilities such as transitioning subsystems to contingency cooling, powering down nonessential loads, and executing pump recovery actions. As individual crew members interact with their assigned subsystems, the VSM continuously evaluates the impact of these actions on overall system stability and timeline constraints. If progress on one responsibility is delayed or produces unexpected results, the VSM can proactively reallocate tasks, adjust action sequences, or propose contingency pathways, allowing the crew to converge on a solution collaboratively and efficiently. In this way, the IDT and VSM pair internalizes the distributed reasoning and coordination functions that are currently performed by ground teams, enabling timely, coherent decision-making despite communication delays.

To further demonstrate the validity of our proposed framework we discuss briefly a case-study based on the capstone demonstration performed by Sherman et al. (2024). In their demonstration, they implemented the distributed hierarchical architecture with ECLSS DT integration. In this scenario an in-transit crew performs a system readiness check of the target habitat before docking. During this check, the crew identifies an existing anomaly and works with the DT to resolve the issue. The crew member responsible for leading the anomaly response process interacted with an adaptive autonomous system that reacts to the operators TWSA based on real-time psychophysiological measures. In their demonstration the TWSA models were localized within an autonomous agent aboard the in-transit crew and were not actually twinned. Meaning that, while the system was able to react to the user in real-time, i) the interaction was not part of the DT system, ii) insights gained from the interaction between the two agents were not maintained, iii) the system (control volume mission-level) is unable to predict, plan, and improve task performance without significant effort to collect post task data, iv) the onboard VSM had no access to information like interaction quality, coordination effort, or team performance. Consequently, the system could not adapt mission operations in real time to unexpected events the way NASA’s ground control teams and flight crews collaboratively do under today’s mission-operations paradigm Dempsey (2018).

If the team had utilized the proposed framework, the system would still have been capable of providing real-time support, as the physiological measures remain available. However, through the integration of an IDT, various scenario-based simulations could have been executed that leverage historical interaction data to improve prediction accuracy. For instance, rather than querying the DT solely for what-if scenarios related to the ECLSS, the autonomous agent’s supervisory algorithm could also query the IDT to estimate the most likely outcomes when different levels of autonomy are applied during a task. In addition, the data provenance enabled by the framework would enable the VSM to perform supplementary analyses aimed at enhancing collaboration quality and crew support. For example, prior to interaction with the autonomous agent, the VSM could query a crew member’s individual DT to estimate the astronaut’s psychological state and proactively recommend an appropriate level of autonomy. Similarly, it could query the IDT to assess trust dynamics for that crew member and implement preemptive mitigation strategies to prevent a degradation of trust.

More broadly, current approaches for supporting mission operations rely either on ground-based human expertise (as in current ISS operations) or on onboard automation, with most proposed approaches for future exploration missions focusing on individual autonomous systems or agents (Dempsey, 2018; Valinia et al., 2023; O’hagan and Crocker, 2011). In current operations, interaction-level reasoning, coordination, and performance assessment are performed by flight directors and controllers; however, these capabilities are inherently ground-dependent and therefore unavailable under long communication delays (Dempsey, 2018; Valinia et al., 2023; Drake, 2009a). Existing onboard autonomy approaches, by contrast, operate at the level of individual systems or agents and do not maintain persistent, integrative models of crew–crew or crew–system interactions that reach beyond their control volume, often limiting their ability to reason about collaboration quality, team dynamics, team characteristics, or team performance as operations evolve. This limitation becomes particularly critical during unexpected and dynamically evolving situations, where maintaining effective and efficient operations, while ensuring crew and system safety, depends on the ability to anticipate, evaluate, and adapt to effects arising from crew-crew and crew-system interactions in real time. The proposed IDT framework complements these approaches by enabling onboard systems and crew to make timely, context-appropriate decisions during mission operations, allowing functions traditionally supported by ground control to be (partially) instantiated within the habitat.

While these insights remain qualitative at this stage, they demonstrate that the proposed framework conceptually satisfies the human behavioral and performance requirements defined in Section 2.1.1.

IDT maturity and gaps

As an evolution of conventional digital twin architectures, the proposed IDT framework inherits many of the well-known challenges associated with DT development and deployment. Issues like varying levels of fidelity within the DT leading to complex uncertainty quantification, human interpretability of DT inference results, model repository management, or the distinction between observed, inferred, and predicted system states during decision-making processes are active challenges for the IDT development. However, the IDT introduces additional challenges related to twinning interactions among crew and systems, particularly the development of teaming and collaboration models that generalize across mission contexts. Although this is an ongoing research focus within the human performance modeling community, it is amplified in the context of IDTs, which must accommodate diverse crew configurations and mission demands without extensive training data or continual expert tuning to stay true to the fundamental idea of an IDT.

An additional challenge lies in supporting the integration and interpretation of heterogeneous data sources, particularly when encountering novel situations, rare events, or less common user profiles for which limited prior data exist. In the context of human deep space missions, this challenge is compounded by the inherently limited availability of historical astronaut data and by the fact that operational data collected during missions may also remain sparse or non-representative, as crews and systems operate for extended periods in novel and evolving environments.

Beyond determining how results should be presented, it remains an open question how continuous monitoring of interactions specifically influence trust, interaction patterns, and long-term psychological health (behavioral and social health).

Finally, studies Hwang et al. (2024); Hwang et al. (2023) already exploring the need for adaptive system reconfiguration during long-duration missions need to be further complimented by supporting research areas such as, updating system ontologies, incorporating newly observed interaction patterns, and accommodating previously unanticipated human–system relationships as missions evolve.

In this paper, we propose an Interaction Digital Twin framework designed to capture the unique and emergent constructs that arise when two entities (at least one being human) interact. This novel framework enhances crew autonomy by improving collaboration quality, enabling effective crew–habitat system teaming, and establishing an information management infrastructure to support crew health monitoring. The IDTs presented in this work are grounded in established digital twin principles, incorporating both information and simulation models, as well as a software interface responsible for managing incoming data, queries, and responses. Although this framework is developed to address the specific and unprecedented challenges associated with future deep-space human exploration missions, its principles are broadly applicable to any system in which humans and autonomous agents are expected to collaborate.