Real-Time Digital Twin Systems

Updated 28 December 2025

Real-time digital twins are cyber-physical systems that continuously synchronize virtual models with physical assets using live sensor data and closed-loop control.
Architectural paradigms integrate sensor layers, edge/fog computing, and cloud orchestration to achieve latencies as low as 5–100 ms essential for timely decision-making.
Hybrid modeling approaches combine physics-based methods, machine learning, and PINNs to deliver rapid state estimation, fault detection, and optimized control.

A real-time digital twin (DT) is a cyber-physical system in which a virtual, computational model of an asset, process, or environment maintains a continuously synchronized state with its physical counterpart using live sensor data, closed-loop control, and online inference. Distinguished from batch-mode or offline twins by strict latency and update-rate requirements, real-time DTs drive decision and control in domains ranging from manufacturing, infrastructure, and energy to healthcare and communications, and increasingly rely on hybrid physical–machine-learning surrogates to balance fidelity with computational responsiveness (Hartmann, 2023, Liu et al., 15 Dec 2025, Mohammad-Djafari, 27 Feb 2025, Iraola et al., 12 Jun 2025, Alkhateeb et al., 2023, Knebel et al., 2020, Srinivasan et al., 17 Oct 2024, Olayemi et al., 2 Jun 2024, Adreani et al., 2023, Shu et al., 2022, Quintanilla et al., 4 Jul 2024, Hossain et al., 17 Oct 2024, Zhang et al., 21 Dec 2025).

1. Architectural Paradigms and Core Real-Time Constraints

The key architectural distinction in real-time digital twins is a closed, low-latency feedback loop integrating sensor acquisition, data transfer, model execution, prediction or optimization, and actuator or decision feedback—all bounded such that physical behaviors are tracked or influenced without perceivable lag.

Representative architectures exhibit the following layered structure:

Physical/Process Layer: Real asset, system, or process (e.g., CNC mill (Liu et al., 15 Dec 2025), AP-1000 reactor (Hossain et al., 17 Oct 2024), supercomputer cluster (Bergeron et al., 1 Oct 2024)).
Sensor/Acquisition Layer: High-frequency, heterogeneous sensors (e.g., AE, force, LiDAR, temperature, video, vibration) deliver streaming data via fieldbus, MQTT, or industrial protocols.
Edge/Fog Layer: Edge processors perform low-latency computation (preprocessing, feature extraction, partial simulation, anomaly detection) and can close fast safety or regulation loops (<10 ms) (Knebel et al., 2020, Hartmann, 2023, Iraola et al., 12 Jun 2025). Cloud/fog architectures partition tasks by urgency and bandwidth.
Dataflow and Orchestration: Message brokers (Kafka, RabbitMQ, MQTT), stream analytics, and microservices ensure scalable, event-driven data and command propagation.
Modeling & Digital Twin Layer: Virtual models (physics-based, ML, hybrid) receive live data, perform state estimation, forecast, or optimization.
Control/Feedback Layer: Model-driven decisions are rapidly dispatched back as control actions, recommendations, or operator guidance (Zhang et al., 21 Dec 2025).

Latency budgets reported range from 5–100 ms (machine/process control (Liu et al., 15 Dec 2025, Hartmann, 2023, Mohammad-Djafari, 27 Feb 2025, Knebel et al., 2020)) to 1–2 s (large-scale monitoring and visualization (Bergeron et al., 1 Oct 2024, Adreani et al., 2023)). Edge/fog co-location is essential for millisecond-class latencies, while hybrid edge–cloud–HPC architectures enable scaling and analytics (Iraola et al., 12 Jun 2025, Hartmann, 2023, Hossain et al., 17 Oct 2024, Quintanilla et al., 4 Jul 2024).

2. Modeling Approaches: Hybrid, Multiphysics, and ML Surrogacy

Real-time digital twins blend physics-based ("first-principles") models, data-driven machine learning, and hybrid (e.g., physics-informed neural networks, PINNs) surrogates to meet the dual goals of physical interpretability and computational speed (Hartmann, 2023, Mohammad-Djafari, 27 Feb 2025, Liu et al., 15 Dec 2025, Hossain et al., 17 Oct 2024).

State-space and PDE/PDE-Replacement: Classical dynamical models (ODEs, PDEs) are used when tractable. For example, reduced models of structural health monitoring (Torzoni et al., 2023) or low-order non-linear acoustics (Nóvoa et al., 29 Apr 2024).
Machine Learning Models: Deep operator networks (DeepONet) for high-dimensional field prediction (Hossain et al., 17 Oct 2024), reservoir computing for real-time bias estimation (Nóvoa et al., 29 Apr 2024), LSTM/GRU modules for temporal forecasting (Mohammad-Djafari, 27 Feb 2025), or custom multi-layer perceptrons (MLPs) for tool–work contact (Liu et al., 15 Dec 2025).
Physics-Informed Neural Networks (PINNs): Embed constraints from governing PDEs directly into loss functions, enforcing residual minimization alongside data fidelity—enabling ML surrogacy where data are sparse or physics is complex (Mohammad-Djafari, 27 Feb 2025).
Hybrid MPC: NN-based surrogates (e.g. TiDE) are embedded in model predictive control, allowing multi-step, nonlinear optimization in additive manufacturing within sub-second loops (Chen et al., 10 Jan 2025).
Online Calibration and Self-Tuning: Kalman filtering, ensemble Bayesian methods, and online (meta-)learning adapt model parameters, account for drift, and recover from disturbance (Hartmann, 2023, Nóvoa et al., 29 Apr 2024, Mohammad-Djafari, 27 Feb 2025).

Real-time digital twins operate under computational and update-frequency constraints, necessitating low-order surrogates or highly optimized inference engines; for instance, the AI-driven milling twin achieves <1 ms model inference (Liu et al., 15 Dec 2025), and DeepONet supports ~0.1 s 3D field inference at 10 Hz (Hossain et al., 17 Oct 2024).

3. Data Pipelines, Synchronization, and Latency Engineering

Robust end-to-end synchronization between the physical and virtual worlds is critical. Typical dataflow is:

Sensing: High-frequency data acquisition (up to 100 kHz for AE in milling (Liu et al., 15 Dec 2025)) with hardware buffering to avoid loss.
Streaming: Protocols such as MQTT, Kafka, and OPC-UA enable highly reliable, low-jitter data movement with sub-millisecond variation (Liu et al., 15 Dec 2025, Knebel et al., 2020, Mohammad-Djafari, 27 Feb 2025).
Processing and Feature Extraction: At the edge or fog, raw sensor data are filtered, down-sampled, or feature-engineered (e.g., AE peak amplitude, time–frequency features).
Inference: Surrogate models are executed either at the edge (for strict latency) or in distributed servers/HPC for heavier analytics (Iraola et al., 12 Jun 2025).
End-to-End Counters: Latency is decomposed and budgeted explicitly, e.g. $L_{e2e} = t_{acq} + t_{stream} + t_{proc} + t_{infer} + t_{ctrl}$ , with evidence of 3–5× improvements by tightly pipelined, optimized streaming (10 ms total round-trip in AI-milling DT (Liu et al., 15 Dec 2025); 66 ms fog-only in general twin (Knebel et al., 2020)).

Dynamic scheduling and offloading (as in HP2C-DT) allow tasks to be mapped "just-in-time" to edge, cloud, or HPC layers as dictated by deadline and computational load (Iraola et al., 12 Jun 2025).

4. Algorithms for Real-Time Estimation, Fault Detection, and Decision-Making

Digital twins continuously perform estimation, forecasting, and (in higher levels) closed-loop control or operational decision-making:

State Estimation: Extended/Unscented Kalman filters and ensemble methods assimilate fresh data for state tracking and parameter adaptation (Nóvoa et al., 29 Apr 2024, Alkhateeb et al., 2023, Mohammad-Djafari, 27 Feb 2025, Torzoni et al., 2023, Liu et al., 15 Dec 2025, Quintanilla et al., 4 Jul 2024).
Anomaly and Disturbance Detection: Automatic modules (e.g. in SAG mill twin (Quintanilla et al., 4 Jul 2024)) flag outliers and initiate retraining or adaptation to maintain model validity.
Optimization and Predictive Control: Model Predictive Control (MPC) with neural surrogates for proactive process adjustment (Chen et al., 10 Jan 2025), or multi-policy discrete-event simulation for online scheduling (Zhang et al., 21 Dec 2025).
Fault Diagnostics and RUL: Bayesian updating, classification nets, and stochastic degradation models estimate remaining useful life (RUL), fault likelihood, or recommend maintenance (Torzoni et al., 2023, Mohammad-Djafari, 27 Feb 2025).
Human-in-the-Loop: Digital twins can enable online, interactive retraining of RL agents with human demonstration for adaptability and safety (Olayemi et al., 2 Jun 2024).
What-If Simulation: For complex environments (e.g., HPC resource scheduling (Zhang et al., 21 Dec 2025), urban traffic (Adreani et al., 2023), crowd management (Srinivasan et al., 17 Oct 2024)), twins can “fast-forward” under multiple scenarios, ranking candidate decisions within a strict delay budget.

5. Applications Across Domains and Benchmarked Performance

Real-time digital twin systems are pervasive across engineering and process domains, with notable benchmarks:

Domain	Twin Functionality	Update/Latency	Modeling Approach	Performance Highlights
Manufacturing	Milling, additive, SAG mill	10–100 ms	ML surrogates, state-space, RNN, MPC	99.86% accuracy @ 10 ms (milling (Liu et al., 15 Dec 2025)); sub-0.3 s MPC (DED (Chen et al., 10 Jan 2025))
Nuclear/CFD	Full-field virtual sensing	~0.1 s (10 Hz)	DeepONet operator networks	1400× CFD speedup, 2×10⁻² Rel-L2 error (Hossain et al., 17 Oct 2024)
Structural Health	Damage diagnosis, decision loop	<10 ms	DL classifier + dynamic Bayes net	93% classification, real-time control (Torzoni et al., 2023)
Smart City	Traffic, pollution, event replay	<0.5 s	Graph-analytics, ARIMA/LSTM, PDE	20k msg/s ingest, 30 FPS 3D UI (Adreani et al., 2023)
Supercomputing/HPC	System & user monitoring	1–2 s	Preprocessing → real-time 3D Unity	~60 FPS with 2000+ nodes (Bergeron et al., 1 Oct 2024)
Crowd/Airport Dynamics	Crowd flow, infection mitigation	<200 ms	Social-force ODE + UKF	4 cm RMSE, sub-167 ms error correction (Srinivasan et al., 17 Oct 2024)
Scheduling	Adaptive policy selection	<2–3 s	Parallel trace-based what-if sim	11.4% performance gain over static baselines (Zhang et al., 21 Dec 2025)

Other domains include precision surgery (Shu et al., 2022), reinforcement learning for autonomous vehicles (Olayemi et al., 2 Jun 2024, Ali et al., 29 Jan 2025), and process optimization in chemical reactors (Mohammad-Djafari, 27 Feb 2025).

6. Technical and Research Challenges

Key research challenges for real-time DTs include:

Scalability: Managing DTs spanning thousands of entities, billions of data points, or exascale events (e.g., urban DT (Adreani et al., 2023, Alkhateeb et al., 2023)).
Distributed/Heterogeneous Compute: Exploiting edge–cloud–HPC hierarchies to dynamically allocate tasks by latency and computational intensity (Iraola et al., 12 Jun 2025, Hartmann, 2023).
Uncertainty Quantification & Robustness: Online QC, error estimation, and resilience to sensor failures and cyberattack are increasingly integral (Hartmann, 2023, Mohammad-Djafari, 27 Feb 2025, Liu et al., 15 Dec 2025).
Model Adaptivity: Continual learning, transfer/adaptation, and domain-bridging to handle concept drift and changing operational envelopes (Hossain et al., 17 Oct 2024, Ali et al., 29 Jan 2025).
Standardization and Interoperability: APIs (OPC-UA, NGSI), semantics, and data models are essential for integration with production IIoT and control systems (Hartmann, 2023, Mohammad-Djafari, 27 Feb 2025, Adreani et al., 2023).
Human–DT Collaboration: Contextualized operator interfaces, AR overlays, natural language querying, and mixed-reality feedback loop completion (Bergeron et al., 1 Oct 2024, Shu et al., 2022, Liu et al., 15 Dec 2025).
Research Directions: Advanced PINNs for multi-physics, federated/differential privacy, uncertainty-aware scheduling, zero-copy data movement, and semantic web integration (Hartmann, 2023, Liu et al., 15 Dec 2025, Iraola et al., 12 Jun 2025, Srinivasan et al., 17 Oct 2024).

7. Evaluation, Validation, and Open Benchmarks

Performance evaluation is multi-dimensional:

Latency and Throughput: Sub-10 ms round-trip in manufacturing/robotics (Liu et al., 15 Dec 2025, Ali et al., 29 Jan 2025); 1–2 s for cluster-scale monitoring (Bergeron et al., 1 Oct 2024).
Accuracy: ML surrogates report 99+% test set accuracy or <5% estimation error (Chen et al., 10 Jan 2025, Liu et al., 15 Dec 2025, Mohammad-Djafari, 27 Feb 2025); field validation via root-mean-square error, mean absolute error, or decision success rate.
Resource Utilization: Profiling indicates that edge devices maintain <50% CPU utilization for streaming + inference pipelines, while GPU-accelerated surrogates enable sub-second inference on high-D grids (Hossain et al., 17 Oct 2024).
Control or Economic Impact: Use cases demonstrate quantitative gains in throughput, downtime reduction, process yield, scheduling makespan, and device availability (Zhang et al., 21 Dec 2025, Adreani et al., 2023, Hartmann, 2023, Mohammad-Djafari, 27 Feb 2025).

Scalability and robustness are validated against increasing sensor populations, message rates, or fault/event bursts, and by evaluating fallback or recovery under network delays or sensor loss (Knebel et al., 2020, Adreani et al., 2023, Cakir et al., 19 Aug 2024).

Real-time digital twins, as realized across fields, are characterized by strict cyber-physical synchronization, latency-aware distributed architectures, hybrid modeling, and adaptive, low-overhead inference and decision workflows. These systems span application domains from precision manufacturing to city-scale digital infrastructure, setting new benchmarks for the orchestration of data, models, and control in complex, dynamic environments (Hartmann, 2023, Liu et al., 15 Dec 2025, Mohammad-Djafari, 27 Feb 2025, Alkhateeb et al., 2023, Iraola et al., 12 Jun 2025, Hossain et al., 17 Oct 2024, Knebel et al., 2020, Adreani et al., 2023, Bergeron et al., 1 Oct 2024, Shu et al., 2022, Chen et al., 10 Jan 2025, Zhang et al., 21 Dec 2025, Nóvoa et al., 29 Apr 2024, Srinivasan et al., 17 Oct 2024, Quintanilla et al., 4 Jul 2024, Olayemi et al., 2 Jun 2024, Ali et al., 29 Jan 2025, Cakir et al., 19 Aug 2024, Torzoni et al., 2023).