Real-Time Digital Twins

• Real-time digital twins are executable virtual replicas that continuously ingest live sensor data to update models with millisecond to second latency.
• They integrate multi-layered architectures including sensing, edge preprocessing, low-latency communication middleware, and AI/physics-based model engines for closed-loop operations.
• Applications span advanced manufacturing, smart infrastructure, and power grids, demonstrating sub-10 ms end-to-end update cycles and enhanced predictive control.

Real-time digital twins (DTs) are executable, continuously coupled virtual counterparts of physical systems that ingest live sensor data, update internal models at tight latency constraints (typically millisecond–second scales), and, when in closed-loop operation, issue actuation or control commands. These systems are foundational to Industry 4.0, advanced manufacturing, cyber-physical systems, and smart infrastructure, where real-time synchronization, predictive analytics, and adaptive control must operate under communication, compute, and reliability constraints.

1. Fundamental Principles and System Architectures

The architecture of a real-time digital twin comprises tightly integrated sensing, inference, actuation, and communication layers, forming a closed loop with the physical system. Core structural elements generally include:

• Sensor and Data Acquisition Modules: Multi-modal data streams (e.g., force, vibration, AE, position, temperature) are sampled, typically at high frequency (e.g., 100 kHz for machining, sub-second or faster for smart grids or urban sensing) (Liu et al., 15 Dec 2025).
• Edge Preprocessing: Real-time feature extraction, denoising, and windowing (e.g., 0.1 s windows, double buffering to prevent loss).
• Communication Middleware: Low-latency streaming protocols such as MQTT, OPC UA, Kafka, and sometimes CoAP or WebSocket, orchestrate data from edge to DT model (Liu et al., 15 Dec 2025, Knebel et al., 2020).
• AI/Physics-Based Model Engine: Physics-based solvers (FEA/FEM, reduced-order PDE models), ML/DL predictors (MLPs, CNNs, RNNs, transformers, PINNs, DeepONets), or hybrid/physics-informed surrogates constitute the core inference capability, updated per streaming batch.
• Feedback and Control Engine: Responsible for threshold-based logic, PID or MPC, and real-time actuation via G-code, PLC, or fieldbus interfaces.
• Visualization/UI: Real-time dashboards, GUIs, and data integration tools (e.g., RESTful APIs, FastAPI, Unity/VR, Grafana, Plotly), with feedback rates matching or exceeding the sensing window (Liu et al., 15 Dec 2025, Stadtmann et al., 2024).

A key architectural pattern is the hierarchical or distributed computing continuum, in which "edge," "cloud," and often "HPC" layers are orchestrated for optimal latency and compute throughput (Iraola et al., 12 Jun 2025, Knebel et al., 2020).

Data flow and scheduling is quantified as

$T_{e2e} = t_{acq} + t_{comm} + t_{proc} + t_{ctrl}$

with update rate constraint $f_{ctrl} = 1 / T_{e2e}$. For advanced industrial DTs, $T_{e2e} \leq 10\,\mathrm{ms}$ is achievable and required for high-performance closed-loop operation (Liu et al., 15 Dec 2025).

2. Data Streaming, Middleware, and Synchronization Protocols

The backbone of real-time digital twins is fast, deterministic data movement and synchronization across system tiers.

• Messaging Protocols: MQTT and OPC UA dominate at the device/gateway level due to their lightweight, pub/sub and semantic features, while Kafka and Redpanda handle high-throughput, ordered streaming between edge/cloud and archival tiers (Liu et al., 15 Dec 2025, Iraola et al., 12 Jun 2025).
• Latency Management: Techniques include edge-side windowing, double buffering, and message aggregation (e.g., phasor averaging in power systems for 10× bandwidth reduction) (Iraola et al., 12 Jun 2025).
• Dynamic Offloading and Scheduling: Functions (control, simulation, analytics) are dynamically routed to edge, cloud, or HPC resources based on a cost function that weights compute and round-trip communication time, typically tuned via task urgency parameters $\alpha$, $\beta$:

$$\text{select}(r) = \arg\min_{r\in\{\text{edge},\text{cloud},\text{HPC}\}} \left( \alpha T_r^{compute} + \beta T_r^{comm} \right)$$

• Synchronization Mechanisms: Edge "shadow" buffers, windowed processing, and efficient API/REST endpoints for multi-client VR or web clients ensure low-latency, consistent views (Knebel et al., 2020, Stadtmann et al., 2024).
• Performance Benchmarks: In power grids, end-to-end latencies of $<10\,\mathrm{ms}$ are achieved for local edge functions; batch HPC simulations scale with hundreds of nodes at strong scaling efficiency $E(N)>0.9$ (Iraola et al., 12 Jun 2025).

Fog computing architectures may interpose an intermediate layer to further reduce response times by half or more, meeting sub-200 ms constraints in IoT-heavy deployments (Knebel et al., 2020).

3. Virtual Modeling: Physics-Based, Data-Driven, and Hybrid Approaches

Real-time DTs rely on a spectrum of modeling techniques:

Physics-Based Models

• Full-order FEM/FDM solvers (Maxwell, Navier–Stokes, vibration PDEs) provide highest fidelity but are computationally intensive, often unsuitable for millisecond updates (Liu et al., 15 Dec 2025).
• Model Order Reduction (MOR): Methods such as Proper Orthogonal Decomposition (POD), DEIM, Krylov subspace reduction yield surrogate models matching target error thresholds ($\leq1\%$ relative error) at 20–100× speed-up, enabling sub-second or faster update cycles (Bonari et al., 2024, Cherifi et al., 2022).
• Lumped-parameter surrogates (e.g., equivalent circuits, thermal RC networks) are used for ultra-fast estimation (e.g., 50 ms updates) (Cherifi et al., 2022).

Data-Driven and AI/ML Surrogates

• Supervised ML: SVMs, decision trees, MLPs, CNNs, transformers, and RNNs/LSTMs, trained with features from high-rate sensing; neural operators such as DeepONet accelerate CFD and reactor-field predictions by $10^3$–$10^4\times$ vs. classical simulations (Liu et al., 15 Dec 2025, Hossain et al., 2024).
• Hybrid Integrations: Physics-informed neural networks (PINNs) and physics-guided loss functions,

$L = L_{data} + \lambda L_{physics}$

incorporate domain constraints, enhancing generalization and sample efficiency (Liu et al., 15 Dec 2025, Mohammad-Djafari, 27 Feb 2025).

• Quantum-Classical Surrogates: Hybrid QMLP architectures, leveraging SPD-based embeddings and PQCs, deliver $10^8$–$10^{10}\times$ lower inverse FE error for structural DTs, albeit with current limitations due to hardware (Alavi et al., 30 Jul 2025).

Training and Quantitative Benchmarks

• Training splits: typically 80/20 train/test, cross-validated for robustness (Liu et al., 15 Dec 2025, Hossain et al., 2024).
• Binary cross-entropy or MSE used for loss; early stopping and Adam optimizer prevalent.
• Application-specific results: 99.86% test accuracy for milling contact status in sub-10 ms loop (Liu et al., 15 Dec 2025); relative L2 errors ≪0.1 for DeepONet-based reactor surrogates at 1400× acceleration (Hossain et al., 2024).

4. Low-Latency Strategies and Real-Time Performance Metrics

Meeting strict latency and real-time constraints across diverse physical domains demands:

• Edge Inference: Pushing inference close to the machine (edge PLCs, embedded GPUs/CPUs) achieves sub-10 ms decision rates for manufacturing and network DTs (Liu et al., 15 Dec 2025, Iraola et al., 12 Jun 2025).
• Model Compression: Pruning, quantization, and parallel streaming (Kafka Streams, Spark Streaming) reduce computation and serialization overhead, supporting > 10,000 windows/s rates (Liu et al., 15 Dec 2025).
• Data Aggregation: Rolling windows and domain-specific aggregation reduce communication needs by up to 90% (Iraola et al., 12 Jun 2025).
• Benchmark Comparisons: Physics-based DTs update in seconds–minutes; ML cloud DTs in 50–100 ms; edge AI DTs attain <10 ms (typ. $T_{e2e} \approx 10\,\mathrm{ms}$, jitter $<$10 ms) (Liu et al., 15 Dec 2025).
• Scenario-Specific Metrics:
  • Urban simulation DTs: Model order reduced wind-solver achieves $\approx$0.1–0.5 s per update, 20–100× faster than FOM (Bonari et al., 2024).
  • Network twins: Adaptive PID achieves tracking MAE reduction by 45% and halves settling time (2 s) for live traffic synchronization (Sengendo et al., 23 Oct 2025).
  • Industrial drives: End-to-end latency $\sim$300 ms with error bounds ≤5% on ultimate KPIs (Cherifi et al., 2022).

5. Case Studies and Domain-Specific Implementations

Manufacturing – Extreme-Low-Latency Milling DT

A sensorized CNC milling system samples AE at 100 kHz, extracts features in 0.1 s windows, and utilizes an MLP ([1–16–16–8–1]) classifier for tool-work contact at 99.86% accuracy. Edge computing and double-buffering yield a total response time $\approx$10 ms, meeting high-throughput production requirements (Liu et al., 15 Dec 2025).

Structural Health and Civil Infrastructure

In large-scale bridge monitoring, a hybrid quantum-classical surrogate predicts full-field nodal displacements from low-dim. sensors in $<$40 ms, with $>10^8\times$ error reduction versus classical MLP, supporting real-time SHM cycles (Alavi et al., 30 Jul 2025).

Wireless Communications & Network Control

Real-time network twins integrate ray tracing, ML, and state-prediction policies (e.g., DRL) for beamforming and resource management, with sub-20 ms or even sub-ms loops on GPU, supporting 6G/URLLC scenarios (Alkhateeb et al., 2023, Zhu et al., 21 May 2025). Scenario-adaptive PID-in-the-loop delivers robust state alignment across dynamic wireless topologies (Sengendo et al., 23 Oct 2025).

Urban and Smart Infrastructure

Urban DTs for contaminant dispersion leverage fully automated pipeline from 2D/3D OSM input to reduced-order CFD solves and GIS-mapped guidance, supporting real-time emergency decision-making with cycle times of 0.1–0.5 s (Bonari et al., 2024).

Power Grids and HPC-Driven Systems

HP2C-DT offloads analytic and simulation workloads dynamically across edge, cloud, and HPC nodes, ensuring sub-10 ms for urgent control while enabling hour-scale data generation via near-ideal strong scaling at cluster scale (Iraola et al., 12 Jun 2025).

6. Scalability, Fault-Tolerance, and Human–DT Interaction

• Scalability: Microservices and stateless orchestrators (e.g., COMPSs, Kapacitor, Docker Compose) allow for live scaling across nodes, handles hundreds of clients, and enables real-time UI interaction (Iraola et al., 12 Jun 2025, Cakir et al., 2024, Adreani et al., 2023).
• Fault Tolerance: Fallback mechanisms (e.g., data reconstitution from last-known-good state, ETL chain robustness) maintain $P_{avail} > 0.999$ under moderate network/API failures (Cakir et al., 2024).
• Visualization: Real-time 3D, VR, and web dashboards (Unity, Deck.gl, FusionLayer) provide visualization of real-time and predicted states, what-if scenario analysis, and support for user-driven decision logic (Stadtmann et al., 2024, Adreani et al., 2023).
• Edge–Cloud/Edge–HPC Partitioning: Design guidelines dictate strict partitioning of latency-sensitive loops to edge, with computationally heavy analytics/batch processes on cloud/HPC (Iraola et al., 12 Jun 2025, Hartmann, 2023).
• Model Adaptation: Online transfer learning, continuous (edge/cloud) monitoring of model drift, and streaming re-training of AI surrogates ensure fidelity as regimes evolve (Liu et al., 15 Dec 2025, Hossain et al., 2024).

7. Limitations and Open Research Directions

Research frontiers in real-time digital twins involve:

• Ultra-low-latency Networking: Leveraging 5G/6G and ultra-reliable low-latency protocols for sub-millisecond loops.
• Autonomous DTs: Online transfer learning, adaptive self-evolving AI models, and robust uncertainty-quantified predictions in nonstationary or adversarial environments (Liu et al., 15 Dec 2025).
• Semantic Interoperability: Developing common standards and ontologies (OPC UA, DDS profiles) to allow plug-and-play across domains and vendors (Liu et al., 15 Dec 2025).
• Cybersecurity: Architectures for trust, resilience to adversarial sensor/actuator streams, and blockchain-based model/data provenance are recognized as necessary safeguards (Hartmann, 2023).
• Human–DT Interaction: Explainable AI interfaces, mixed-initiative decision support, and operator-in-the-loop modes are essential for mission-critical and regulated environments.
• Computational Efficiency: Research into distributed/parallel MOR, hardware-aware ML, and quantum/hybrid acceleration for further reduction in latency and energy per update.
• Bespoke Domain Extensions: Extension of operator networks, hybrid quantum surrogates, and multi-fidelity DTs to new physical domains and multi-agent/federated systems.

Major challenges include full bidirectional integration with physical assets (actuation as well as sensing), maintaining fidelity at massive scale, semantic interoperability, and dynamic partitioning of workloads under variable compute and network resources (Liu et al., 15 Dec 2025, Sengendo et al., 23 Oct 2025, Iraola et al., 12 Jun 2025, Alkhateeb et al., 2023).

---

References

• Real-Time AI-Driven Milling Digital Twin Towards Extreme Low-Latency (Liu et al., 15 Dec 2025)
• Hybrid Quantum Classical Surrogate for Real Time Inverse Finite Element Modeling in Digital Twins (Alavi et al., 30 Jul 2025)
• HP2C-DT: High-Precision High-Performance Computer-enabled Digital Twin (Iraola et al., 12 Jun 2025)
• Towards Real-Time Urban Physics Simulations with Digital Twins (Bonari et al., 2024)
• Hierarchical modeling for an industrial implementation of a Digital Twin for electrical drives (Cherifi et al., 2022)
• AI-empowered Real-Time Line-of-Sight Identification via Network Digital Twins (Zhu et al., 21 May 2025)
• Real-Time Digital Twins: Vision and Research Directions for 6G and Beyond (Alkhateeb et al., 2023)
• Virtual Sensing-Enabled Digital Twin Framework for Real-Time Monitoring of Nuclear Systems Leveraging Deep Neural Operators (Hossain et al., 2024)
• A Cloud-Fog Computing Architecture for Real-Time Digital Twins (Knebel et al., 2020)
• Smart City Digital Twin Framework for Real-Time Multi-Data Integration and Wide Public Distribution (Adreani et al., 2023)
• Data Integration Framework for Virtual Reality Enabled Digital Twins (Stadtmann et al., 2024)
• State Aware Traffic Generation for Real-Time Network Digital Twins (Koktas et al., 16 Sep 2025)
• Building Network Digital Twins Part II: Real-Time Adaptive PID for Enhanced State Synchronization (Sengendo et al., 23 Oct 2025)