Neural Digital Twins

• Neural Digital Twins are virtual replicas that fuse physics-informed models with neural network surrogates to predict, control, and adapt to complex system behaviors.
• They combine techniques like model reduction, neural ODEs, and probabilistic uncertainty quantification to achieve real-time inference and robust calibration.
• Applications across manufacturing, biomedicine, and quantum technologies demonstrate significant speedups and enhanced predictive accuracy.

A Neural Digital Twin (NDT) is a virtualized, dynamically evolving computational replica of a physical or biological system, whose generative core is realized by neural networks and hybrid AI surrogates. Unlike conventional digital twins based purely on physics solvers or classical data-driven regression, NDTs unify physics-informed (or physically constrained) neural network parameterizations, real-time synchronization to streaming sensor or measurement data, and, crucially, end-to-end uncertainty quantification, control, or inference capabilities. The NDT paradigm encompasses model classes ranging from reduced-order surrogate ODEs/PDEs and probabilistic graphical models to large-scale hybrid simulators with closed-loop, feedback-driven adaptation, and is realized across domains from online manufacturing, biomedicine, and materials, to neuroscience, quantum technologies, and industrial systems.

1. Mathematical and Architectural Foundations

Neural Digital Twins are typically structured around a reduced-order or hybrid surrogate dynamical system:

• Physics-informed model reduction: The high-fidelity system, for instance a PDE of the form ∂ₜu = 𝒩(u), is projected onto a low-rank (e.g., POD or spectral) basis {φ₁, ..., φ_r}, yielding coefficients a(t) (Ning et al., 1 Nov 2025).
• Neural surrogate dynamics: Rather than directly discretizing the physics right-hand side, a neural ODE approximates the projected evolution: ȧ(t) = f_θ(a(t)), where f_θ is a neural network parameterized by θ.
• Hybrid fusion: In frameworks like PhysiNet, the physics-based and neural model outputs are linearly or nonlinearly fused via trainable mixing weights: $$\hat y(x) = w_{\rm phys} q_{\rm phys}(x) + w_{\rm nn} q_{\rm nn}(x)$$ (Sun et al., 2021). The weights are adapted online as more data become available.
• Probabilistic closure and UQ: For uncertainty quantification, Gaussian filtering or moment closure is applied to propagate (μ(t), P(t))—the mean and covariance of the reduced/latent state—by linearizing the neural flow around μ and integrating ODEs for (μ, P), analogous to an extended Kalman filter (Ning et al., 1 Nov 2025).

Neural operator-based twins for spatially extended systems adopt spectral decompositions (e.g., Laplacian eigenfunctions) and parameterize unknown nonlinear operators via neural nets, as in LENO for biomarker dynamics on cortical surfaces (Xu et al., 23 Jun 2026).

2. Training, Calibration, and Uncertainty Quantification

• Data regime: Training samples are obtained from high-fidelity simulations, physical experiments, or historical process data. For multi-parametric surrogates, a hyper-network h_ψ(t, μ) delivers time/parameter-conditioned statistics for Bayesian weight sampling (Ning et al., 1 Nov 2025); for biomarker dynamics, PET/fMRI projection onto spectral modes yields (β^n) time-series (Xu et al., 23 Jun 2026).
• Loss functions: Training minimizes tasks-dependent losses:
  • Physics/data hybrid loss: ℒ = ℒ_data + λ ℒ_phys, with ℒ_data enforcing proximity to observed values (sensor readings) and ℒ_phys imposing governing ODE/PDE residuals at collocation points (Mohammad-Djafari, 27 Feb 2025).
  • Variational / Bayesian inference: In variational surrogates, parameters governing meso-structural features are given Gaussian priors, and optimization seeks to maximize evidence or minimize negative log-likelihood of observed responses (Robertson et al., 19 Dec 2025).
  • Maximum mean evidence lower bound (ELBO): Used in co-learning of offline and online neural weights, ensuring that the approximate posterior of time-local weights matches the true posterior given the data (Ning et al., 1 Nov 2025).
• Uncertainty propagation: Analytical Taylor expansions or Jacobian-based approximations are used to propagate input/sample uncertainty through the neural surrogate (e.g., VDMN analytic mean and covariance prediction) (Robertson et al., 19 Dec 2025).

3. Real-Time Inference, Feedback, and Control

• Ensemble-free UQ: Moment closure methods propagate mean and covariance via closed ODE systems, completely bypassing the need for costly pathwise ensembles. Filtered Neural Galerkin moment equations yield speedups >10× compared to ensemble runs for uncertainty propagation (Ning et al., 1 Nov 2025).
• Calibration via neural surrogates: Attentive neural processes (ANP) conditioned on context data act as surrogates for expensive calibration objectives, scaling to high-dimensional parameter spaces and enabling batch-parallel Bayesian optimization for near-real-time model inversion (Chakrabarty et al., 2021).
• Feedback integration: Multi-scale and feedback-centric architectures employ state-space representations at each biological or physical scale, with continuous Kalman filtering or ensemble Bayesian updates synchronizing the virtual and physical states (e.g., for neural/biomedical twins) (Zhang, 22 Jun 2026, Bina et al., 4 Jan 2026). Formal feedback contracts specify exchanged state variables, latency/reliability requirements, and decision authority.
• Closed-loop and optimal control: PDE-constrained treatment optimization for dynamical systems (e.g., biomarker regulation in AD) solves forward–backward sweep algorithms for the ODE representation of biomarker spectral coefficients under time-parameterized interventions (Xu et al., 23 Jun 2026). In medical twins, neural-network controllers are trained by differentiating through the twin (ODE, ABM, or hybrid), directly optimizing patient-specific or population-representative outcomes (Böttcher et al., 2024).

4. Domain Applications and Quantitative Benchmarks

NDTs have been instantiated in high-impact domains, with demonstrable real-world and computational gains:

Domain	Example NDT Frameworks	Quantitative Metrics
Uncertainty-aware surrogates	Filtered Neural Galerkin, VDMN	>10× speedup over ensemble; Quantile agreement with 50-member ensemble in 1 run; NRMSE <1% fields (Ning et al., 1 Nov 2025, Robertson et al., 19 Dec 2025)
Biomedical control	ABM + Neural ODE controller	Steady-state parameter error within 1σ of optimum; Robustness to stochasticity; Outperforms classical ODE baselines (Böttcher et al., 2024)
Neuroinformatics/neuroscience	Hierarchical feedback, brain “Digital Twin”	>90% parameter recovery in calibration, simulated BOLD–fMRI correlation >0.75, suicide prediction R=0.655 (Zhang, 22 Jun 2026, Lu et al., 2023)
Operator learning for disease	Laplacian eigen-operator neural twin (AD)	Spectral, nodal accuracies 87–88% (amyloid), 81–82% (tau); Treatment reduces biomarker burden by 30–50% (Xu et al., 23 Jun 2026)
Reduced FEEC battery models	FEEC–Transformer twin (Whitney Forms)	Exact conservation, ∼0.1s full-field battery inference, speedup ≃3.1×10^8× over LES (Kinch et al., 9 Aug 2025)

Domain-specific instantiations emphasize modularity, data efficiency, and physically-constrained expressivity, such as the Deep Material Network for materials UQ (Robertson et al., 19 Dec 2025), batch Bayesian calibration in building physics (Chakrabarty et al., 2021), and fast calibration-driven control of additive manufacturing (Kannapinn et al., 2024).

5. Implementation Complexity, Efficiency, and Modularity

Prominent NDT frameworks achieve practical real-time fidelity through:

• Reduced order and locality: POD-, FEEC-, or spectral decomposition compress the physics state to a tractable r∼5−100, enabling sub-millisecond or sub-second inference on commodity GPUs (Ning et al., 1 Nov 2025, Kinch et al., 9 Aug 2025).
• Analytic and autodiff-based differentiation: For mean/covariance propagation or inverse uncertainty quantification, derivatives/Jacobians are computed via autodiff engines (e.g., Torch functorch) or analytic expressions (Robertson et al., 19 Dec 2025).
• Hardware acceleration: Analogue neural ODE solvers implemented in memristive hardware realize in-memory computation, yielding 10–100× reductions in both energy and latency compared to digital ODE solvers (Chen et al., 2024).
• Modularity and integration: Pre-trained surrogates can be deployed in multi-agent simulations, digital twin VR environments, or optimization/control loops without invasive remeshing or solver modification (Kinch et al., 9 Aug 2025, Gunaratne et al., 29 Mar 2025).

6. Limitations, Open Questions, and Future Research

Current limitations and active research issues in neural digital twins include:

• Scalability and generalization: Efficient surrogates for 10^5–10^10 DOF, and robust interpolation/extrapolation across geometry, material, or parameter space, demand further methodological advances (Ning et al., 1 Nov 2025, Kannapinn et al., 2024).
• Feedback contract formalization: Questions of interoperability, latency/SLA guarantees, and authority partitioning (human-in-the-loop, safety validation) remain outstanding for multi-scale deployments (Zhang, 22 Jun 2026).
• Data sufficiency and bias: Training surrogates in scarce or biased-data regimes requires further fusion of mechanistic priors and transfer learning schemes; the quality of training data fundamentally limits performance (Robertson et al., 19 Dec 2025).
• Uncertainty quantification (UQ): Most existing frameworks focus on Gaussian/first-moment closures; capturing heavy-tailed, multi-modal, or adversarial uncertainty remains a challenge (Ning et al., 1 Nov 2025).
• Real-time adaptation and lifelong learning: On-line adaptation of weights, structure, and privacy-compliant learning to accommodate drift, non-stationarity, and cross-twin sharing/federated learning are active areas (Bina et al., 4 Jan 2026).

7. Significance and Outlook

Neural Digital Twins provide a mathematically rigorous foundation for uncertainty-aware, feedback-capable, controllable, and interpretable digital replicas of high-dimensional, multi-physics, or biological systems. Substantial computational accelerations have already enabled real-time predictive control, risk quantification, and adaptive inference in fields spanning rare materials, additive manufacturing, industrial automation, precision medicine, brain–machine interfaces, and quantum devices. Deep integration of neural surrogates with physical constraints, model reduction, and measurement feedback distinguishes NDTs from both pure black-box ML models and classical digital twins.

Their ongoing evolution is poised to underlie the next generation of accountable, interoperable, and trustworthy digital systems for both science and engineering (Ning et al., 1 Nov 2025, Robertson et al., 19 Dec 2025, Zhang, 22 Jun 2026, Xu et al., 23 Jun 2026, Kinch et al., 9 Aug 2025).