Mixed Digital Twin Architectures
- Mixed digital twin is a family of architectures that integrates physical and virtual spaces with heterogeneous models and synchronized live data.
- These systems employ decomposition and specialized sub-twins or multi-layer stacks to coordinate diverse representations and maintain semantic alignment.
- Applications span autonomous vehicles, clinical systems, and industrial IoT, demonstrating enhanced operational insights and actionable feedback.
Searching arXiv for papers on mixed digital twins and closely related hybrid digital twin architectures. Mixed digital twin denotes a family of digital-twin architectures in which the twin is not treated as a single isolated virtual replica, but as a coupled system that mixes physical and virtual spaces, heterogeneous model classes, or specialized operational twins within a shared feedback loop. In the cited literature, the term is used for cloud-mediated mixed spaces in which physical and virtual entities coexist and interact, for hybrid twins that combine physics-based and data-driven models, for semantically aligned multi-layer model stacks, and for clinical systems that tightly synchronize digital state, physical instrumentation, and mixed-reality or haptic interfaces (Dong et al., 2022, Kunzer et al., 2022, Abbasi et al., 17 Dec 2025, Ping et al., 15 Mar 2026). Across these variants, the recurring properties are live data coupling, synchronization across representations, and operational outputs that can inform or actuate the physical system.
1. Terminological scope and principal meanings
The literature does not present a single canonical definition of mixed digital twin. Instead, several closely related meanings recur. In vehicle-road-cloud research, mixed digital twin, often abbreviated mixedDT, is defined by a three-space architecture: physical space, virtual space, and mixed space, with the mixed space integrating physical and virtual entities so that they can coexist and interact in real time (Dong et al., 2022). In hybrid modeling and maintenance research, the mixed or hybrid twin combines machine learning with physics-based modeling, or more generally combines multiple model forms so that the strengths of one compensate for the weaknesses of another (Kunzer et al., 2022). In model-driven digital twins, “mixed” refers to interoperability across data, models, metamodels/schemas, and ontologies, with explicit mechanisms for maintaining semantic coherence across abstraction layers (Abbasi et al., 17 Dec 2025). In clinical and training systems, the term is applied to synchronized digital–physical environments in which the twin aligns real anatomy, devices, or procedures with mixed reality, haptics, or physical phantoms (Wang et al., 16 May 2025, Ping et al., 15 Mar 2026, Shu et al., 2022).
| Usage in the literature | Defining combination | Representative papers |
|---|---|---|
| Three-space mixedDT | Physical space + virtual space + mixed space | (Dong et al., 2022, Dong et al., 18 Mar 2026) |
| Action-interactive mixedDT | Bidirectional action-level interaction between physical and virtual entities | (Dong et al., 18 Mar 2026) |
| Hybrid model twin | Physics-based models + machine learning or data-driven components | (Kunzer et al., 2022, Azangoo et al., 2021) |
| Multi-layer semantic twin | Data, models, metamodels, ontologies, and alignment mechanisms | (Abbasi et al., 17 Dec 2025) |
| Digital–physical clinical twin | Synchronized digital models + physical phantoms, MR, or haptics | (Wang et al., 16 May 2025, Ping et al., 15 Mar 2026, Shu et al., 2022) |
Despite this plurality, the survey literature still provides a baseline digital-twin scaffold. A minimally viable framework comprises seven elements: Physical Twin Asset, Digital Twin, Instrumentation, Analysis, Digital Thread, Live Data, and Actionable Information (Kunzer et al., 2022). Mixed digital twins inherit these elements, but they add heterogeneity in representation, interaction, or control.
2. Architectural patterns
A central architectural pattern is decomposition rather than monolithic modeling. The oncology operations framework explicitly rejects a single all-purpose twin and instead proposes a mixed, multi-twin architecture in which specialized twins collaborate around a shared Cancer Care Path. The three explicitly named twins are the Medical Necessity Twin, Care Navigator Twin, and Clinical History Twin, embedded in a layered stack consisting of a data layer, knowledge layer, retrieval and prompting layer, reasoning layer, agent orchestration layer, and operational output layer (Pandey et al., 2024). This is mixed in at least four senses at once: specialized sub-twins, symbolic knowledge, LLM-based reasoning, and workflow software logic.
A second pattern is the explicit separation of physical and digital sides within one cyber-physical system. The gantry-crane exemplar divides the system into a physical side—the lab-scale crane, its controllers, and operator interface—and a digital side containing CAD models, a kinematic model, a trajectory generation service, simulation, historical logging, visualization, and continuous validation (Mertens et al., 17 Jul 2025). The architecture is described using the C4 model, emphasizing containers and interactions rather than an undifferentiated twin.
A third pattern is the three-space cloud-mediated architecture used in mixed vehicle systems. MCCT and MSH-MCCT organize experimentation into physical, virtual, and mixed platforms, with the mixed platform running in the cloud and functioning as the integration, synchronization, and control hub (Dong et al., 2022, Dong et al., 18 Mar 2026). In this design, physical and virtual entities do not interact directly; they interact through cloud fusion, state aggregation, and command dispatch.
A fourth pattern is the multi-layer model stack. In heterogeneous model alignment, the digital twin is framed as a model-driven system with four design-phase layers—data, models, metamodels/schemas, and ontologies—whose relations must remain synchronized through lifecycle evolution (Abbasi et al., 17 Dec 2025). This suggests that mixed digital twin is often less a single model type than a systems-integration pattern that coordinates representations at different semantic and operational levels.
3. Interaction, synchronization, and validation
Mixed digital twins are distinguished from static digital models by closed-loop synchronization. In the gantry-crane exemplar, the Logging Application continuously stores crane state in TimescaleDB; the Grafana dashboard visualizes real and simulated traces; the Trajectory Generator Service synthesizes optimal movements; and the Continuous Validation Service compares measured and simulated trajectories against a threshold and alerts operators when divergence becomes excessive (Mertens et al., 17 Jul 2025). The twin therefore has memory, prediction, and runtime self-checking.
In clinical systems, synchronization requirements become spatiotemporal and sensorimotor. Twin-S uses an Atracsys FusionTrack 500 optical tracker, calibrated drill, phantom, and stereo microscope models, and real-time physics simulation to update drilled anatomy at 28 FPS with an overall runtime of 35.7 ms per frame; the reported overall average drilled-region error is 1.39 mm with standard deviation 0.62 mm (Shu et al., 2022). Surgi-HDTMR adds metrically co-registered MR overlays and a depth-adaptive haptic model; hand increments stream at 90 Hz, the haptic servo loop runs at 1 kHz, rendering occurs at 90 Hz, and end-to-end latency is kept below one 90 Hz frame, approximately 11.1 ms (Ping et al., 15 Mar 2026). These systems illustrate a stronger requirement than mere data freshness: visual, geometric, and force consistency must remain phase-aligned.
The vehicle-infrastructure literature extends synchronization from data to action. IMPACT introduces an L5 “Interactable” level above L4 “Optimizable”, defined as “Actuating the physical via the virtual (Actionable Interaction),” with “bidirectional action-level interaction between physical and virtual entities, requiring direct human intervention in their control.” The cloud performs latency compensation, spatio-temporal alignment, coordinate unification, instruction validation, and mapping to entity-specific interfaces; measured one-way delays have Gaussian means roughly from 0.36 ms to 4.23 ms, with 99th percentiles up to 8.23 ms (Dong et al., 18 Mar 2026). The shift is from synchronized observation to synchronized intervention.
In semantically grounded twins, validation includes ontology-level coherence. The alignment framework combines flexible conformance through JSMF with SSM-OM: Semantics and Structure-aware Metamodel Ontology Matching, where candidate matches are generated from SBERT embeddings, Jaro–Winkler lexical similarity, and Jaccard structural overlap, then re-ranked by an LLM that classifies relations as exactMatch, closeMatch, or broadMatch (Abbasi et al., 17 Dec 2025). Here, synchronization is not only temporal or geometric; it is also semantic.
4. Domain-specific realizations
In oncology clinical operations, the mixed digital twin is primarily an operational and administrative intelligence system rather than a tumor-biophysics simulator. The framework is designed to reconcile fragmented EHR data, NCCN evidence-based guidelines, and payer prior authorization / medical necessity criteria. The Medical Necessity Twin evaluates approval likelihood under guideline and payer policies; the Care Navigator Twin recommends the next diagnostic or therapeutic step; and the Clinical History Twin constructs a longitudinal timeline from structured and unstructured EHR content. The implementation mechanisms explicitly include RAG over a vector database, ICL, CoT prompting, function calling, and the integration of symbolic logic for deterministic tasks such as prior-authorization simulation. Prior work cited for the medical-necessity component reports 86.2% accuracy for item-level prediction and 95.6% for overall checklist judgment (Pandey et al., 2024).
In surgery and intervention, mixed digital twins are used to close perception–action loops. Twin-S continuously updates a CT-derived skull-base phantom model during drilling and projects segmentation masks and depth-aware overlays into the microscope view (Shu et al., 2022). Surgi-HDTMR couples a benchtop microsurgical robot, Quest 3-based MR, and a digital twin that drives a depth-adaptive haptic model with parameters for stiffness, damping, viscosity, friction, puncture thresholds, and adhesion (Ping et al., 15 Mar 2026). In cardiovascular intervention, a patient-specific dynamic digital-physical twin is built from 4D-CTA, a transparent silicone coronary phantom, pulsatile flow simulation, virtual angiography, and binocular stereo-vision guidewire tracking. Reported outcomes include 80.9% morphological consistency between virtual and real angiography, Dice similarity coefficients of 0.741–0.812 for guidewire motion, and mean trajectory errors below 1.1 mm (Wang et al., 16 May 2025).
In connected and autonomous vehicle research, mixed digital twins are used as hybrid safety and validation environments. MCCT integrates a 9 m × 5 m sand-table physical platform at 1:14 scale, a Unity virtual platform, and a cloud-based mixed platform supporting Microsoft HoloLens, driving simulators, and multi-vehicle platooning (Dong et al., 2022). MSH-MCCT extends this to multi-source human-in-the-loop testing with three Logitech G29 simulators and one InnoSimulation high-fidelity simulator, permitting humans and CAV algorithms to control both physical and virtual vehicles through the cloud; in a safety-critical platooning scenario, three collisions occurred, but all involved at least one virtual vehicle, so no physical damage occurred (Dong et al., 18 Mar 2026). IMPACT and the I-VIT testbed make the same principle explicit at the action level, so that hazardous corner cases can be realized in mixed physical–virtual traffic without requiring full physical realization of every entity (Dong et al., 18 Mar 2026).
In industrial and infrastructure contexts, the mixed digital twin often emphasizes reproducibility, validation, and hybrid modeling. The gantry-crane exemplar is publicly available, includes CAD, simulation, logging, optimization, and continuous validation, and packages its Python components as gantrylib, with database, MQTT broker, and dashboard deployed through Docker and docker compose (Mertens et al., 17 Jul 2025). In process industry, a hybrid twin is formed in Apros by replacing the heater E100 in Tank100 with a Keras / TensorFlow LSTM encoder-decoder model trained on process history data, yielding an as-built brownfield twin that can be continually updated (Azangoo et al., 2021). In plant-factory teleoperation, a mixed-reality interface replaces pre-built 3D farm models with a live camera-derived immersive scene processed using Nerfies, over which virtual control panels and object highlights are embedded for remote monitoring and control (Ban, 2022). In mixed autonomous traffic safety analysis, the twin combines UAV LiDAR, OpenStreetMap, GPS, inclinometer data, SUMO, CARLA, and NVIDIA PhysX; the resulting high-fidelity TTC formulation reduces error relative to traditional TTC, with MAE 0.29 s and RMSE 0.32 s versus 0.71 s and 0.97 s (Zhang et al., 24 Apr 2025).
5. Relation to adjacent concepts and recurrent misconceptions
A recurring misconception is that a mixed digital twin is simply a digital twin with a more elaborate visualization front end. The literature rejects this reduction. In the vehicle-road-cloud papers, mixedDT is not merely an MR display; it is a cloud-based architecture in which the mixed space supports synchronized coexistence, data fusion, and control dispatch across physical and virtual entities (Dong et al., 2022). The plant-factory paper makes the same distinction from a different angle: the goal is remote work and operation through a mixed-reality environment built from real camera scenes and IoT state, “rather than simulation of components” (Ban, 2022).
A second misconception is that every virtual entity in a mixed digital twin must be a strict one-to-one replica of a physical entity. This is explicitly contradicted in the mixedDT literature. Virtual vehicles in MCCT and MSH-MCCT are not necessarily tied to physical vehicles (Dong et al., 18 Mar 2026). IMPACT goes further by defining L5 interactivity precisely through action-level coupling between physical and virtual entities, including cases where the virtual side exists independently and still directly affects the physical scenario (Dong et al., 18 Mar 2026).
A third misconception is that mixed digital twin implies a single broad model. The oncology framework instead uses specialized collaborating twins, and the model-alignment work shows that the twin may itself be a coordinated stack of heterogeneous artifacts rather than a monolithic executable model (Pandey et al., 2024, Abbasi et al., 17 Dec 2025). This makes decomposition, schema evolution, and semantic alignment first-class concerns.
A fourth misconception is that mixed digital twin is synonymous with purely data-driven AI. The maintenance survey argues that hybridization is often necessary because machine learning alone faces rarity-of-failure, labeled-data scarcity, spurious correlations, and black-box limitations, while physics-based models alone can be too slow, assumption-bound, or disconnected from operational data (Kunzer et al., 2022). The oncology system similarly mixes LLM reasoning with symbolic knowledge and workflow logic, rather than relying on generative output alone (Pandey et al., 2024). In this sense, mixed digital twin frequently denotes a principled integration of complementary epistemic regimes.
6. Limitations, open problems, and likely trajectories
The principal limitations vary by subdomain, but several themes recur. Semantic drift and model evolution remain difficult: static mappings and manual updates are described as brittle and error-prone, motivating adaptive conformance and ontology-grounded alignment mechanisms (Abbasi et al., 17 Dec 2025). In brownfield process systems, data collection effort, feature selection, integration complexity, and sampling-rate mismatch complicate the construction of as-built hybrid twins (Azangoo et al., 2021). The maintenance survey additionally emphasizes sensor robustness, missing data, workplace adoption, and security as persistent deployment barriers (Kunzer et al., 2022).
Clinical systems expose fidelity limits sharply. Twin-S remains constrained by optical line of sight; the average drilling depth in the experiments was only 1.57 mm, and future work is directed toward complementing optical tracking with vision-based methods (Shu et al., 2022). The coronary digital–physical twin is validated on a single patient, and the authors note limits in material realism, distal-vessel fidelity, and imaging-modality mismatch (Wang et al., 16 May 2025). The plant-factory MR interface is explicitly not yet commercial-grade; 45° capture intervals were too wide for high-quality radiance-field reconstruction, and real-time streaming still requires further development (Ban, 2022).
Vehicle and infrastructure testbeds raise scalability and interaction-management questions. IMPACT does not yet resolve all multi-request concurrency and leaves open how multimodal LLMs might be used for richer human–machine interaction and corner-case generation (Dong et al., 18 Mar 2026). MSH-MCCT demonstrates multi-driver mixed traffic platooning, but its complexity implies harder synchronization and protocol-management problems as participant counts increase (Dong et al., 18 Mar 2026). The traffic-safety twin built on CARLA, SUMO, and PhysX demonstrates the value of physics-informed TTC, but also makes clear that slope, friction, and mass materially affect safety estimates, so accurate map reconstruction and parameter calibration remain essential (Zhang et al., 24 Apr 2025).
The oncology ecosystem points toward another trajectory: expansion from decision support to broader operational automation. The proposed future tasks include ambient note-taking, patient education, self-care advice, and scheduling, together with self-verification techniques such as self-consistency and fact-checking to reduce hallucinations (Pandey et al., 2024). A plausible implication is that future mixed digital twins will increasingly function as orchestrated operational ecosystems in which perception, knowledge retrieval, symbolic rules, learned models, human interfaces, and physical assets are coupled continuously rather than episodically.
Taken together, the literature portrays mixed digital twin not as a single technology stack or narrow application label, but as an architectural response to a recurring systems problem: real-world operations require simultaneous coordination of physical assets, virtual abstractions, heterogeneous models, and human decision-makers. The most mature formulations therefore emphasize not virtual resemblance alone, but synchronized interoperability, validation against reality, and controlled exchange of data, semantics, and action across the physical–digital boundary.
