Physics-Informed Digital Twin

• Physics-informed digital twins are virtual replicas that integrate governing equations with data-driven models for predictive, calibrated, and real-time system control.
• They employ methodologies such as residual penalties, neural surrogates, and embedded physics blocks to ensure consistency with fundamental laws.
• Applications span maritime dynamics, thermal processes, additive manufacturing, and more, enhancing accuracy and computational efficiency across diverse domains.

to=arxiv_search 彩神争霸大发json {"^^^^^^^^2query2 digital twin\"","max_results":^^^^2all:\2query2^^^^,"sort_by":"submittedDate","sort_order":"descending"} RTLU шәрjson to=arxiv_search 《凤凰大参考json &&&^^^^2query2^^^^&&&)v^^^^2all:\2 \"Authors\": [\"Narges Ahmadi\", \"Binghang Yuan\", \"Leandro Paes\", \"Afsaneh Rabiei\", \"Mohamad Yaghoubi\", \"Mohammad S. Yousefi\", \"Stig Pedersen\"], \"Summary\": \"Maritime operations in polar regions require accurate dynamic models to ensure safe ship transit through ice-infested waters. This paper proposes a ship digital twin framework capable of real-time state estimation and load prediction under sea ice impact. The digital twin combines a first-principles mass-spring-damper model of a vessel with a multi-fidelity correction mechanism based on a Random Forest predictor. The method continuously updates the low-fidelity physical model using high-fidelity sensor and simulation data, enabling accurate state estimation under varying ice conditions. A particle filter is employed to estimate hidden states and uncertain parameters online, while domain randomization is used to improve the generalizability of the correction model. The model is validated on both simulated and laboratory-generated impact scenarios and demonstrates reduced prediction error over baseline methods.\"}, \"Physics-informed Digital Twins for Learning and Control of AGN Feedback in Galaxy Simulations\": {\"arXiv ID\": \"(&&&^^^^2all:\2^^^^&&&)v^^^^2all:\2 \"Authors\": [\"Ugo Liva\", \"Lena Sanders\", \"Ming-Feng Ho\", \"Alena Kataeva\", \"Abhirup Ghosh\", \"Stefano Borgani\", \"Romain Teyssier\", \"Franck H\'enault-Brunet\", \"Andres Villaescusa-Navarro\", \"David A. A. Hanson\", \"Siddhartha Mishra\"], \"Summary\": \"The energy released by active galactic nuclei (AGN) is a key ingredient in the evolution of galaxies and their host clusters. Capturing AGN feedback accurately in galaxy formation simulations remains challenging due to the unresolved small-scale physics and the need for computationally efficient control models. We develop a differentiable, equation-informed neural surrogate as a digital twin for AGN feedback, coupled with a high-resolution simulation-based teacher model. The surrogate combines image-supervision and physical consistency losses to reproduce the teacher's outputs from sparse local observations and supports closed-loop control through gradient-based optimization. We demonstrate that the twin recovers global thermodynamic and morphological observables while enabling policy search over feedback parameters at a fraction of the simulation cost.\"}, \"Physics-Informed Digital Twin for Multiple Input Multiple Output Thermal Process Control\": {\"arXiv ID\": \"(Russo et al., 13 May 2025)v^^^^2all:\2^^^^\", \"Authors\": [\"Vincent Dayanidhi\", \"Shuo Wang\", \"Andres Tovar\"], \"Summary\": \"Conventional thermal process modeling methods that rely solely on first-principles or experimental system identification often struggle to simultaneously achieve both accuracy and computational efficiency, especially in the presence of nonlinear coupling, disturbances, and changing operating conditions. To address these challenges, a physics-informed digital twin (PIDT) framework is developed for monitoring and control of a multiple-input multiple-output (MIMO) thermal process. The framework integrates lumped-capacitance thermodynamic equations with neural correction models to capture system dynamics efficiently. To further enhance performance, sensor placement is optimized using a Fisher information matrix criterion to maximize state observability and estimation quality. The PIDT is combined with a moving horizon estimator for online state estimation and a nonlinear model predictive controller for trajectory tracking. Experimental validation on a Thermal Coupled Distillation Column setup shows that the PIDT significantly improves prediction accuracy and reduces tracking errors compared with baseline gray-box models, while maintaining real-time computational feasibility.\"}, \"A Physics-Informed Neural Network Digital Twin for a Regenerative Cryogenic Refrigerator: Coupling Experimental Data and Thermodynamics\": {\"arXiv ID\": \"(Baillou et al., 14 Mar 2025)v^^^^2all:\2^^^^\", \"Authors\": [\"Pierandrea Vigni\", \"Marco Guida\", \"Emanuele Piacibello\", \"Andrea Ferrero\", \"Michele De Benedetti\", \"Matteo De Pascale\"], \"Summary\": \"The dynamic and non-linear behavior of regenerative cryogenic refrigerators under variable operating conditions presents a challenge for conventional modeling methods, which are often either computationally expensive or fail to capture the underlying thermodynamic mechanisms. To address this, we propose a hybrid digital twin based on a Physics-Informed Neural Network (PINN) trained on experimental data to estimate key state variables and output temperatures while enforcing thermodynamic constraints. The framework combines a reduced-order thermodynamic model with a multi-objective optimization strategy to estimate latent quantities such as pressure amplitudes and phase shifts, requiring only a limited set of measurable temperatures as inputs. The resulting digital twin enables accurate transient prediction and robust parameter estimation across a range of operating conditions, offering a practical tool for real-time monitoring and control.\"}, \"A Physics-Informed Digital Twin Model for Predicting the Thermo-Mechanical Behaviour of Additively Manufactured Parts\": {\"arXiv ID\": \"(Bustos-Espinoza et al., 16 Feb 2025)v^^^^2all:\2^^^^\", \"Authors\": [\"Raúl Morales\", \"Carlos M. Molina-Solana\", \"Angel Arcos-Jiménez\", \"Natividad Martínez Madrid\", \"J. Miguel Corral-Guerrero\", \"Pedro José Arcenegui-Domínguez\", \"María José Molina-Solana\"], \"Summary\": \"Metal additive manufacturing (MAM) introduces complex thermal and mechanical phenomena during the build process that can critically affect part quality and structural integrity. Numerical simulations can capture these phenomena but are often computationally expensive and impractical for real-time process control. This paper presents a Physics-Informed Digital Twin Model (PIDTM) that predicts thermal fields, residual stresses, and deformations in MAM components during fabrication. The approach combines finite element simulations, a recurrent neural network, and explicit thermodynamic and mechanical constraints to build a reduced-order representation that supports in-situ prediction and parameter updates. Experimental validation on laser powder bed fusion specimens demonstrates the model's ability to capture thermal and strain evolution with low latency.\"}, \"Physics-Informed Digital Twin for Predictive Maintenance of Mechanical Systems\": {\"arXiv ID\": \"(Frisk et al., 15 Jan 2025)v^^^^2all:\2^^^^\", \"Authors\": [\"Aaditya Mishra\", \"Jackson M. Collins\", \"Piyush Rai\", \"Suyash Rai\", \"Ayan Biswas\"], \"Summary\": \"Physics-based digital twins have emerged as powerful tools for prognostics and health management, but their practical deployment remains constrained by the high cost of parameter estimation and the limited availability of field data under fault conditions. This paper presents a physics-informed digital twin framework for predictive maintenance of mechanical systems that combines analytical state-space models with deep neural surrogates and Bayesian updating. The method uses a residual learning architecture to estimate health parameters from sparse sensor data and quantify remaining useful life under varying load profiles. Case studies on rotating machinery datasets show improved fault detection lead time and robustness relative to purely data-driven baselines.\"}, \"Physics-Informed Neural Network Digital Twin for Dynamic Tray-Wise Modeling of Distillation Columns under Transient Operating Conditions\": {\"arXiv ID\": \"(Patra et al., 25 Mar 2026)v^^^^2all:\2^^^^\", \"Authors\": [\"Hasan Burak Tokgoz\", \"Atakan Arslan\", \"Korhan Cengiz\", \"Abdurrahman Sabanovic\"], \"Summary\": \"Digital twin technology, when combined with physics-informed machine learning with simulation results of Aspen, offers transformative capabilities for industrial process monitoring, control, and optimization. In this work, the proposed model presents a Physics-Informed Neural Network (PINN) digital twin framework for the dynamic, tray-wise modeling of binary distillation columns operating under transient conditions. The architecture of the proposed model embeds fundamental thermodynamic constraints, including vapor-liquid equilibrium (VLE) described by modified Raoult's law, tray-level mass and energy balances, and the McCabe-Thiele graphical methodology directly into the neural network loss function via physics residual terms. The model is trained and evaluated on a high-fidelity synthetic dataset of 96^^^^2all:\2^^^^ timestamped measurements spanning 8 hours of transient operation, generated in Aspen HYSYS for a binary HX/TX distillation system comprising ^^^^2all:\2^^^^6 sensor streams. An adaptive loss-weighting scheme balances the data fidelity and physics consistency objectives during training. Compared to five data-driven baselines (LSTM, vanilla MLP, GRU, Transformer, DeepONet), the proposed PINN achieves an RMSE of ^^^^2query2^^^^.^^^^2query2query2all:\2 for HX mole fraction prediction (R^2 = ^^^^2query2^^^^.9887), representing a 44.6% reduction over the best data-only baseline, while strictly satisfying thermodynamic constraints. Tray-wise temperature and composition profiles predicted under transient perturbations demonstrate that the digital twin accurately captures column dynamics including feed tray responses, reflux ratio variations, and pressure transients. These results establish the proposed PINN digital twin as a robust foundation for real-time soft sensing, model-predictive control, and anomaly detection in industrial distillation processes.\"}, \"A Modular Digital Twin Architecture for Power Electronics in DC Microgrids\": {\"arXiv ID\": \"(González et al., 16 Feb 2025)v^^^^2all:\2^^^^\", \"Authors\": [\"Matías A. Fiorelli\", \"Manuel N. Acosta\", \"Juan I. Yuz\"], \"Summary\": \"This work proposes a digital twin architecture for power electronics applications in DC microgrids, aimed at improving scalability, maintainability, and interoperability. The architecture is structured in five modular layers: data, model, service, synchronization, and application. To describe and connect these layers, the Asset Administration Shell (AAS) standard is adopted as a unified digital representation framework. Each layer can contain multiple modules, enabling flexible integration of hardware interfaces, physical and data-driven models, communication protocols, and digital twin services such as monitoring and predictive control. The proposed architecture is validated through a bidirectional DC-DC converter case study, where a state-space model of the converter is combined with a Particle Swarm Optimization calibration service and real-time hardware-in-the-loop experiments. Results show reduced synchronization latency and improved maintainability compared to monolithic twin designs.\"}, \"A Digital Twin Framework for Soft and Wearable Robots\": {\"arXiv ID\": \"(Yi et al., 25 Jan 2025)v2\", \"Authors\": [\"Austin Bonnet\", \"Khaled Mahmud\", \"Devjyoti Paul\", \"Onur Ozcan\"], \"Summary\": \"Wearable robots require continuous sensing, adaptive modeling, and timely user feedback to operate safely and effectively. Yet, their nonlinear and user-specific dynamics make accurate online estimation difficult, particularly when measurements are sparse or delayed. This work presents a modular digital twin framework for wearable and soft robotic systems that combines finite-element or reduced-order mechanical models with data-driven estimators and reinforcement-learning-based adaptation. The framework emphasizes synchronization, state estimation, and user-in-the-loop control, and is demonstrated on a soft exosuit for gait assistance with real-time biomechanical feedback.\"}}] A physics-informed digital twin is a virtual representation of a physical system that combines mathematical modeling with data-driven inference so that prediction, calibration, control, and uncertainty quantification remain anchored to governing physics rather than to statistical correlation alone. In the cited literature, this role is realized through several distinct but related mechanisms: PDE or DAE residuals added to the training objective, hard-wired physics blocks embedded directly in network architectures, parameterized numerical solvers optimized from data, and probabilistic discrepancy models that explicitly represent missing physics. The resulting systems span groundwater contamination, coastal flooding, buildings, medical hemodynamics, optical networks, additive manufacturing, industrial distillation, radar sensing, and other domains (Wang et al., 2022).

^^^^2all:\2^^^^. Definition and conceptual scope

In one formulation, a digital twin is “a virtual replica of a real-world physical phenomena that uses mathematical modeling to characterize and simulate its defining features,” while in another it is a framework that maintains continuous synchronization between a physical asset and its virtual counterpart through online parameter estimation under uncertainty (&&&^^^^2all:\2query2^^^^&&&). Physics-informedness refers not to a single algorithmic family but to the imposition of first-principles structure on that twin.

The literature distinguishes several ways of imposing that structure. In PINN-style formulations, governing equations enter as residual penalties such as PRESERVED_PLACEHOLDER_^^^^2query2^^^^, together with boundary, initial-condition, and data terms. In operator-learning formulations, a surrogate learns the solution operator of a PDE while additional losses enforce PDE residuals, derivative consistency, or boundary conditions. In architecture-centric formulations, physics is enforced exactly in the forward model through static “physics blocks,” with residual learning layers correcting only the part not captured by the analytical model. In probabilistic formulations, a low-fidelity simulator is retained explicitly and an additive discrepancy term is learned with a Gaussian process so that model-form error is represented rather than ignored. These alternatives are all presented as physics-informed digital twins in the supplied sources (&&&^^^^2all:\2all:\2^^^^&&&).

A recurring distinction is between hybrid and purely data-driven twins. A hybrid digital twin is described as merging a purely data-driven “digital twin” with a physics-based “virtual twin” to form a gray-box surrogate capable of real-time predictions and reliable extrapolation. This distinction is operational rather than rhetorical: several studies report that physics constraints stabilize extrapolation, reduce data requirements, or support latent-parameter identification in regimes where data-only surrogates deteriorate (&&&^^^^2all:\22^^^^&&&).

2. Mathematical foundations

The governing models used in physics-informed digital twins span ODEs, DAEs, PDEs, and kinematic equations. In groundwater contamination, the twin is built around saturated porous-medium flow and advection–dispersion:

PRESERVED_PLACEHOLDER_^^^^2all:\2^^^^

with Darcy velocity $\mathbf v=-K\nabla h$ (Wang et al., 2022). In coastal flooding, the surrogate targets the nonlinear shallow-water equations for sea-surface height $\eta$ and horizontal velocity $u$, including mass conservation and momentum balance with Coriolis, bottom stress, and atmospheric forcing (&&&^^^^2all:\2^^^^4&&&).

In building digital twins, the physical core is written in DAE form,

$$0=f_{\mathrm{DAE}}(\dot x(t),x(t),u(t),\theta_1), \qquad y(t)=h_{\mathrm{DAE}}(x(t),u(t),\theta_2),$$

and calibration is posed as optimization over constant but unknown parameters $\theta$ (&&&^^^^2all:\2^^^^5&&&). In optical-fiber parameter estimation, the digital twin is anchored directly to the nonlinear Schrödinger equation residual

$$R\bigl(s(z,t);\theta\bigr)=\frac{\partial s}{\partial z}+\frac{\alpha}{2}s+\frac{\mathrm{j}\beta_2}{2}\frac{\partial^2 s}{\partial t^2}-\mathrm{j}\gamma |s|^2 s=0,$$

with $\theta=\{\alpha,\beta_2,\gamma\}$ (&&&^^^^2all:\2^^^^6&&&).

The loss constructions reflect these governing equations. A DD-PINN for incompressible Navier–Stokes uses

$$L(\theta)=L_{\mathrm{res}}+L_{\mathrm{BC}}+L_{\mathrm{IC}}+w_{\mathrm{data}}L_{\mathrm{data}},$$

where PRESERVED_PLACEHOLDER_^^^^2all:\2query2^^^^ is assembled from continuity and momentum residuals, and the data term incorporates sensor or CFD guide points (&&&^^^^2all:\2all:\2^^^^&&&). The groundwater twin uses a more specialized composite loss combining mean-relative error, spatial-derivative error, contaminant-boundary sharpness above MCL, and a no-flow boundary soft penalty, with

PRESERVED_PLACEHOLDER_^^^^2all:\2all:\2^^^^

and PRESERVED_PLACEHOLDER_^^^^2all:\22^^^^ in practice (Wang et al., 2022). In distillation, the total loss is written as

PRESERVED_PLACEHOLDER_^^^^2all:\2^^^^3

with PRESERVED_PLACEHOLDER_^^^^2all:\2^^^^4 composed of modified Raoult’s law, tray-level mass balance, tray-level energy balance, and McCabe–Thiele operating-line constraints (Patra et al., 25 Mar 2026).

This variety suggests that “physics-informed” is best understood as a modeling principle rather than a single training recipe. The principle is that admissible predictions are constrained by equations, constitutive relations, geometry, or conservation laws that remain meaningful outside the training set.

3. Architectural patterns

One major pattern is neural operator surrogacy. The groundwater twin employs U-FNO, which augments each Fourier Neural Operator layer with a U-Net skip-connection structure. The reported variants differ in temporal treatment: U-FNO-3D treats spacePRESERVED_PLACEHOLDER_^^^^2all:\2^^^^5time as a single 3D field and predicts all future timesteps in one shot, whereas U-FNO-2D is recurrent and preserves causality at the cost of accumulated error. The architecture combines low-frequency spectral kernels with a U-Net “convolutional core” for local multi-scale refinement (Wang et al., 2022). The coastal-flood twin similarly learns a mesh- and resolution-invariant solution operator using FNO layers that interleave learned local channel mixing with band-limited spectral convolutions (&&&^^^^2all:\2^^^^4&&&).

A second pattern is PINN and DD-PINN construction. Here the surrogate is a standard neural network, but collocation sampling, residual losses, and boundary enforcement transform it into a physics-guided model. The DD-PINN framework explicitly studies residual-aware, gradient-aware, vorticity-aware, and combined RGV adaptive sampling, extends the input space to include the Reynolds number for parameter-scalable inference, and adds multi-fidelity data terms with fidelity-dependent weights. No architectural change is required for the multi-fidelity version; the difference lies in the structure of the loss (&&&^^^^2all:\2all:\2^^^^&&&). The reinforced-concrete beam example applies the same philosophy to both a temporal ODE surrogate inspired by the harmonic oscillator and a spatial mixed-variable PINN based on linear elasticity (&&&^^^^2all:\22^^^^&&&).

A third pattern is direct embedding of analytical models into the network graph. In the Physics Encoded Residual Neural Network architecture, physics blocks are static differentiable operators and residual blocks are inserted in parallel so that the model learns corrective terms without removing analytical structure. The paper emphasizes that the overall loss remains purely data-driven because the forward model itself already enforces the physical equations exactly (Zia et al., 2024). A related but even more interpretable construction appears in PIDT, where the “neural operator” is replaced by a parameterized split-step Fourier method. The only trainable quantities are the segment-wise physical coefficients of the solver, and a PDE residual is enforced at interior points through a physics-informed loss (&&&^^^^2all:\2^^^^6&&&).

A fourth pattern is inverse or calibration-oriented twin construction. ANP-BBO treats the expensive map PRESERVED_PLACEHOLDER_^^^^2all:\2^^^^6 as the object to be learned by an Attentive Neural Process and then optimized via batch Bayesian optimization with UCB acquisition. The medical hemodynamics framework adopts a two-step, self-supervised structure: first pretrain a differentiable simulator on synthetic ODE-solver outputs, then freeze that simulator and train an inverse network that reconstructs physiological measurements from noninvasive inputs through the simulator itself (&&&^^^^2all:\2^^^^5&&&). This suggests a useful taxonomy of physics-informed digital twins into operator surrogates, residual PINNs, physics-embedded architectures, and differentiable inverse twins (&&&^^^^2all:\2query2^^^^&&&).

4. Calibration, synchronization, and uncertainty

Calibration is central because a digital twin is useful only if it remains aligned with the asset it represents. In the building case, calibration minimizes

PRESERVED_PLACEHOLDER_^^^^2all:\2^^^^7

with PRESERVED_PLACEHOLDER_^^^^2all:\2^^^^8, and ANP-BBO scales this procedure by retraining an ANP surrogate on the accumulated dataset and proposing PRESERVED_PLACEHOLDER_^^^^2all:\2^^^^9 parallel candidates with a UCB rule using $\mathbf v=-K\nabla h$^^^^2query2^^^^ (&&&^^^^2all:\2^^^^5&&&). In optical networks, dynamic updating is even more explicit: new input/output power pairs are assimilated, an inverse-update loss is formed from span-end measurement errors plus an SRS-ODE residual, and a three-step cycle alternates between updating the neural operator, updating physical parameters, and jointly refining both (Song et al., 28 Apr 2025).

Several works formalize synchronization as recursive filtering. The cattle thermoregulation twin combines an ODE-based heat-balance model, a Gaussian process for cow-specific deviations, a behavioral Markov chain, and a Kalman filter that fuses model predictions with real-time sensor data. The output is not limited to core body temperature; it also includes heat stress probability and behavioral state distributions (&&&3^^^^2query2^^^^&&&). In spacecraft parameter estimation, Weighted Flow Matching is coupled to an Unscented Kalman Filter through a “virtual sensor” update, where the WFM-derived mean and covariance of parameters become pseudo-measurements inside the filter (&&&3^^^^2all:\2^^^^&&&).

Uncertainty quantification is handled by several non-equivalent mechanisms. DD-PINNs use ensembles of independently initialized models and report predictive mean, variance, and $\mathbf v=-K\nabla h$^^^^2all:\2^^^^ confidence bands as $\mathbf v=-K\nabla h$2 (&&&^^^^2all:\2all:\2^^^^&&&). The groundwater twin reports that predicted heads and concentrations remain within $\mathbf v=-K\nabla h$3 confidence bands obtained from expensive Monte Carlo when five climate clusters are used to span wet-to-dry scenarios (Wang et al., 2022). Learning between digital twins with low-fidelity models and physics-informed Gaussian processes places a GP prior on the discrepancy term and a Bayesian hierarchical prior across individuals, with the stated conclusions that models not accounting for imperfect physical models are biased and over-confident, while models learning between twins reduce uncertainty without becoming over-confident (Spitieris et al., 2022). In the TEDS heat-exchanger study, the probabilistic MvG-SINDyC surrogate supplies $\mathbf v=-K\nabla h$4 credible intervals and is paired with active learning so that informative trajectories are preferentially queried (Nabila et al., 7 May 2026).

A common misconception is that physics-informedness removes the need for calibration. The cited works point in the opposite direction: they repeatedly present physics as a structural prior that must still be updated through measurements, filtering, optimization, or posterior inference.

5. Representative application domains and reported performance

In environmental systems, the groundwater multi-scale digital twin couples site-scale U-FNO surrogates to continental-scale climate compression. The dataset comprises 664 stochastic Amanzi simulations over DOE’s Savannah River Site spanning ^^^^2all:\2^^^^954–2^^^^2all:\2query2query2 On an 8:^^^^2all:\2^^^^:^^^^2all:\2^^^^ split, U-FNO-3D with physics-informed loss achieves MRE $\mathbf v=-K\nabla h$5–$\mathbf v=-K\nabla h$6 and MSE $\mathbf v=-K\nabla h$7; after ^^^^2all:\2submittedDate2query2^^^^ epochs on an NVIDIA A^^^^2all:\2query2query2^^^^, the reported test values are MRE $\mathbf v=-K\nabla h$8 and MSE $\mathbf v=-K\nabla h$9. The model predicts in milliseconds versus minutes for a full-order Amanzi run, corresponding to a reported $\eta$^^^^2query2^^^^–η\eta^^^^2all:\2^^^^ computational saving (Wang et al., 2022). In coastal flooding, an FNO surrogate trained on NEMO data achieves MSE $\eta$2, SSIM $\eta$3, mean spatial correlation $\eta$4, and an over $\eta$5 speedup relative to NEMO (&&&^^^^2all:\2^^^^4&&&).

In built-environment calibration, the four-story office-floor digital twin uses a ^^^^2all:\22^^^^-parameter search domain and 2^^^^2query2query2^^^^ BBO iterations with batch size $\eta$6. On the final 3 days, the reported coefficient of variation RMSE is below $\eta$7, compared with an ASHRAE guideline of $\eta$8, and the estimated parameters are within $\eta$9 of true values for most parameters (&&&^^^^2all:\2^^^^5&&&). In structural and civil applications, the reinforced-concrete beam PINN extrapolates correctly to ^^^^2all:\2^^^^6 s whereas the purely data-driven temporal model “explodes past 6 s”; the same study identifies $u$^^^^2query2^^^^ against a measured value of $u$^^^^2all:\2^^^^ and reports prediction in $u$2 s for the temporal surrogate (&&&^^^^2all:\22^^^^&&&).

In biomedical settings, the cardiac hemodynamics twin identifies patient-specific parameters from noninvasive echocardiogram videos. Reported end-systolic/diastolic volume performance gives EF MAE values of $u$3 on CAMUS and $u$4 on EchoNet for the physics-informed self-supervised method, alongside unsupervised disease detection and in-silico LVAD trials with average EF increase $u$5–$u$6 across EchoNet/CAMUS post-LVAD (&&&^^^^2all:\2query2^^^^&&&). In optical communication networks, the dynamic-updating twin reports up to $u$7 speedup in prediction versus classical numerical methods and a maximum accuracy improvement of $u$8 dB for performance estimation post-device replacement; in the field trial, channel-power RMSE improves from $u$9 dB to $$0=f_{\mathrm{DAE}}(\dot x(t),x(t),u(t),\theta_1), \qquad y(t)=h_{\mathrm{DAE}}(x(t),u(t),\theta_2),$$^^^^2query2^^^^ dB after updating (Song et al., 28 Apr 2025).

In industrial process systems, the distillation-column PINN twin is trained on 96^^^^2all:\2^^^^ timestamped measurements over 8 hours of transient operation and achieves RMSE $$0=f_{\mathrm{DAE}}(\dot x(t),x(t),u(t),\theta_1), \qquad y(t)=h_{\mathrm{DAE}}(x(t),u(t),\theta_2),$$^^^^2all:\2^^^^ with $$0=f_{\mathrm{DAE}}(\dot x(t),x(t),u(t),\theta_1), \qquad y(t)=h_{\mathrm{DAE}}(x(t),u(t),\theta_2),$$2 for HX mole fraction prediction, a $$0=f_{\mathrm{DAE}}(\dot x(t),x(t),u(t),\theta_1), \qquad y(t)=h_{\mathrm{DAE}}(x(t),u(t),\theta_2),$$3 reduction over the best data-only baseline (Patra et al., 25 Mar 2026). In additive manufacturing, two FNO surrogates trained on 75^^^^2query2^^^^ high-fidelity simulations reach final relative $$0=f_{\mathrm{DAE}}(\dot x(t),x(t),u(t),\theta_1), \qquad y(t)=h_{\mathrm{DAE}}(x(t),u(t),\theta_2),$$4 errors of $$0=f_{\mathrm{DAE}}(\dot x(t),x(t),u(t),\theta_1), \qquad y(t)=h_{\mathrm{DAE}}(x(t),u(t),\theta_2),$$5 and $$0=f_{\mathrm{DAE}}(\dot x(t),x(t),u(t),\theta_1), \qquad y(t)=h_{\mathrm{DAE}}(x(t),u(t),\theta_2),$$6 on test data, while closed-loop control reduces surface roughness by approximately $$0=f_{\mathrm{DAE}}(\dot x(t),x(t),u(t),\theta_1), \qquad y(t)=h_{\mathrm{DAE}}(x(t),u(t),\theta_2),$$7–$$0=f_{\mathrm{DAE}}(\dot x(t),x(t),u(t),\theta_1), \qquad y(t)=h_{\mathrm{DAE}}(x(t),u(t),\theta_2),$$8 and keeps the per-update cost below $$0=f_{\mathrm{DAE}}(\dot x(t),x(t),u(t),\theta_1), \qquad y(t)=h_{\mathrm{DAE}}(x(t),u(t),\theta_2),$$9 s on a single GPU (Liu et al., 2024). In sim-to-real radar perception, a physics-informed FMCW twin with calibrated domain randomization yields approximately $\theta$^^^^2query2^^^^ balanced accuracy for binary occupancy detection and approximately $\theta$^^^^2all:\2^^^^ for $\theta$2-person counting using only synthetic training data plus a small unlabeled real calibration set (Trinh et al., 25 Jan 2026).

These examples do not define a single performance frontier, because the tasks differ substantially. They do show that the term “physics-informed digital twin” is used for systems that are simultaneously predictive, computationally compressed, and operationally updateable.

6. Methodological issues, misconceptions, and reported directions

One misconception is that physics-informed digital twins are synonymous with PINNs. The sources explicitly show otherwise. Physics may appear as residual losses in PINNs and neural operators, as exact analytical blocks in residual architectures, as low-fidelity solvers plus GP discrepancy, as interpretable model libraries selected by optimal trees, or as parameterized numerical propagators optimized end-to-end (Zia et al., 2024). A second misconception is that adding physics automatically solves generalization and identifiability. Several papers instead stress practical balancing issues: the reinforced-concrete study states that the choice of loss weights is crucial and that no universal rule exists; ANP-BBO notes that the penalization radius $\theta$3 and UCB weight $\theta$4 may require tuning; PIDT shows that training SSFM only on output-observation loss can recover segment-wise dispersion profiles that deviate significantly from the true constant profile (&&&^^^^2all:\22^^^^&&&).

The limitations reported across domains are also heterogeneous. Some are structural, such as the brainbot twin’s restriction to purely two-dimensional rigid-body kinematics with no explicit leg compliance, frictional forces, or multi-bot interactions (Mammadli et al., 1 Nov 2025). Some arise from observability, as in medical twins where patient imaging covers only partial observables and some parameters remain weakly identified (&&&^^^^2all:\2query2^^^^&&&). Some are computational or statistical, such as the linear growth of ANP retraining cost with data size, or the possibility that ANP Gaussian assumptions under-represent heavy-tailed noise (&&&^^^^2all:\2^^^^5&&&). Others concern simplified physical constitutive models, such as isotropic linear elasticity without explicit crack modeling in civil structures (&&&^^^^2all:\22^^^^&&&).

Reported future directions are correspondingly diverse. The groundwater study names adaptive mesh refinement, inverse-problem calibration against sensor data, and on-the-fly climate-cluster updates as next steps (Wang et al., 2022). The optical-network work frames greenfield instance-specialization followed by brownfield real-time drift compensation as a general lifecycle-management blueprint for other networked physics systems (Song et al., 28 Apr 2025). The medical twin proposes amortized uncertainty quantification, extension to full spatiotemporal PDE-based simulators, and joint training with semi-supervised PDE residuals (&&&^^^^2all:\2query2^^^^&&&). This suggests that the field is moving less toward a single canonical architecture than toward a family of interoperable design patterns for building living, updateable surrogates whose admissible behavior remains constrained by domain physics.