ABSTRACT Physics‐informed neural networks (PINNs) provide a promising paradigm for integrating governing physical laws with data‐driven learning in digital twin systems. However, existing studies and reviews have primarily emphasised offline prediction accuracy or domain‐specific applications, whereas the structural role of PINNs in real‐time model updating and continuous synchronisation remains insufficiently clarified. This review focuses on real‐time updating mechanisms in evolving digital twins and examines how PINN‐based and operator‐learning‐based models support synchronisation between physical assets and their digital counterparts. First, the functional requirements of real‐time digital twin updating are analysed, including online calibration, data–model synchronisation, multitimescale response and application‐dependent latency constraints. Classical and advanced real‐time estimation methods are also compared to clarify their trade‐offs in nonlinearity handling, uncertainty quantification, computational cost and real‐time applicability. Second, the technical mechanisms of PINNs are reviewed from the perspectives of physics‐constrained parameter identification, time‐dependent learning, data–physics fusion, adaptive training and neural operator‐based generalisation. Third, a functional layered framework is proposed to organise PINN‐driven digital twins into interacting components responsible for physical consistency enforcement, synchronisation and data assimilation, online updating and computational acceleration. The methodological roles of PINNs are further clarified by distinguishing their functions as physics‐constrained surrogate models, inverse identification tools and data–physics fusion frameworks. Finally, structural challenges are critically synthesised, including optimisation instability, uncertainty propagation, sparse sensing and identifiability, synchronisation constraints, computational scalability and robustness under distribution shifts. Rather than treating PINNs as general‐purpose prediction models, this review positions them as structural mechanisms for regulating consistency, adaptability and computational efficiency in real‐time digital twin architectures.
ABSTRACT Rapid AI advancements have primarily focused on enhancing digital twins (DTs) using artificial intelligence (AI), as seen in the author's earlier research addressing scaling challenges such as computing resource efficiency and manual insight generation. Conversely, existing research on using DTs to enhance AI remains nascent, mostly limited to DTs serving as data factories for synthetic training and providing on‐demand validation environments to accelerate AI model testing and training. This work proposes that DTs have the potential to enhance AI through other diverse applications and explores a mutualistic symbiosis where DTs can improve AI via context, memory, factual grounding and bridging the ‘trust gap’ through better explainability, demonstrating that AI and DTs act as essential drivers for each other's scalability and development. We demonstrate how DTs facilitate the transition of AI from stateless chatbots to persistent adaptive cognitive agents. Ultimately, the integration of these fields leads to the emergence of cognitive digital twins (CDTs), where this work identifies a potential opportunity where DTs begin to enter the realm of AI, potentially contributing to the advancement of emerging AI concepts such as world models (WMs), obscuring the traditional distinctions between the AI and DTs.
ABSTRACT High‐quality multiobjective engineering optimisation requires large candidate populations to approximate dense Pareto fronts, yet CPU‐bound algorithms such as NSGA‐II become impractical for large populations. Although GPU hardware can relax this constraint, only algorithms whose internal structure permits data‐parallel execution can fully exploit it. We present a PyTorch‐native GPU MOPSO framework that extends the Coello Coello external‐archive design with linearly decaying inertia, boundary‐safe iterative archive pruning and periodic particle mutation, and we empirically characterise how swarm size, GPU throughput, Pareto‐front quality and run‐to‐run reproducibility interact across a range of engineering optimisation problems. The principal empirical finding is that GPU hardware makes it practical to run swarms of N=2000–5000 particles and that these larger swarms produce substantially better engineering outcomes: GPU PSO‐5000 achieves 40% lower generational distance than CPU NSGA‐II‐100 on combustion engine calibration (p<0.05, Wilcoxon signed‐rank test, 20 independent seeds), whereas GPU PSO‐2000 achieves the best hypervolume on a 30‐dimensional neural‐network surrogate in 2.5 s, which is 55% faster than CPU NSGA‐II and has 8.3× lower run‐to‐run variance. A scalability study from N=500 to N=20,000 identifies a GPU throughput sweet spot at N=2000 (2.5 μs per particle per iteration), where variance collapses 12× relative to N=500, enabling reproducible single‐run deployment without ensemble averaging. These gains arise from two compounding effects: GPU batch inference reduces per‐particle surrogate evaluation cost by 16× over CPU, whereas PSO’s parallel update rule avoids the ON2 nondominated sorting bottleneck that prevents NSGA‐II from scaling beyond N≈200. A controlled GPU‐vs‐GPU comparison isolates this second effect, revealing a 74× timing gap at N=2000 attributable solely to algorithmic structure. Results generalise across a real‐world UCI Gas Turbine dataset, a 6D/4‐objective engine calibration problem and a fixed 60‐s wall‐clock budget protocol under which GPU PSO’s 5‐s convergence enables 12 independent seeds within a single NSGA‐II run budget.
ABSTRACT For large‐aircraft surface inspection, nonuniform viewpoint distribution and wide task scope make Euclidean‐distance clustering difficult to jointly achieve path efficiency and workload balance. To address this, we propose a density‐adaptive clustering distance dDA, allowing density information to influence cluster‐boundary adjustment throughout the iterative assignment process. Based on a C919 digital‐twin model, a nonuniform viewpoint set satisfying coverage and viewing‐angle constraints is generated; then, a Gaussian‐kernel local density on the viewpoint set (KDE‐style) and an S‐type mapping are used to construct an adaptive weight, and k‐means++ algorithm with dDA is used to complete multi‐unmanned aerial vehicles (UAVs) task allocation under total‐distance minimisation, coefficient of variation (CV)‐controlled workload balance and single‐UAV energy constraints. For fair comparison, all methods share a unified ‘Nearest‐Neighbour + 2‐opt’ backend to evaluate each UAV closed‐loop length. Simulations on the C919 digital geometric model with 582 viewpoints and 3 UAVs show that DAF–CPP achieves the shortest mean total path (679.1 ± 2.6 m), the lowest mean maximum path (239.5 ± 13.4 m) and the lowest mean CV (0.1025 ± 0.0166) among eight allocation baselines under a unified 20‐seed protocol. Extended fairness metrics and Holm‐corrected tests further support the statistical significance of the allocation improvement.
ABSTRACT With the growing penetration of renewable energy, traditional power flow methods struggle to capture the multidimensional uncertainties and abrupt meteorological shifts in active distribution networks. To address this, the paper proposes a digital twin‐driven probabilistic power flow framework based on the semi‐invariant method and Gram–Charlier series expansion. In the modelling stage, real‐time data continuously calibrates Weibull, Beta and normal distribution parameters, transitioning grid evaluation to agile online tracking. A highly efficient convolution‐free engine utilises linear matrix mapping to directly derive nodal voltage semi‐invariants, whereas the Gram–Charlier expansion dynamically reconstructs continuous probability distributions. Simulation results on the IEEE 33‐bus system demonstrate that the proposed framework effectively quantifies the spatial accumulation of network uncertainties and the temporal evolution of operational risks, providing grid operators with dynamic probability boundaries of nodal voltage.
