Integrating Machine Learning and Uncertainty Quantification for Real-Time Decision-Making in Digital Twins: A System-Centric Review and Perspective
Organizations: Department of Mechanical Engineering, Northwestern University, 2145 Sheridan Rd., Evanston, 60208, Illinois, USA · Ingram School of Engineering, Texas State University, 601 University Dr., San Marcos, 78666, Texas, USA
Abstract
Digital Twins are evolving from passive virtual representations into closed-loop systems that continuously synchronize with physical assets and drive real-time decisions. Yet predictive modeling, machine learning (ML), uncertainty quantification (UQ), and decision-making have been studied as separate capabilities, leaving limited understanding of their integration and maintenance during online operation. This narrative review, scoped by the functional requirements of real-time decision-making rather than by Digital Twin terminology alone, examines how such decisions are generated and sustained in engineering Digital Twins. We first examine foundational decision-making architectures and representative engineering applications, followed by methods that address their limitations across the decision loop. We identify optimization-based, policy-based, and knowledge-based approaches as the major pathways from virtual representation to action. Across them, ML plays three roles in decision-making: enabling efficient surrogate modeling, transforming heterogeneous physical feedback into decision-relevant information, and accelerating or directly generating decisions. UQ supports uncertainty-aware decisions and model adaptation, while continual validation, drift detection, and adaptation help maintain reliable operation as the physical system evolves. Together, these findings underscore the need for a shift from a model-centric to a system-centric perspective on Digital Twins, where its performance depends on the interactions with sensing, state estimation, computation, decision-making, and adaptation rather than on their standalone performance. Therefore, future research should shift from isolated algorithm development toward system-level co-design, scalable decision-making, and approaches to design, and data-efficient continual adaptation and validation for AI-enabled autonomous systems.
Figures & tables
| Deployment | State | Decision | Stage cost | Role of the Digital Twin |
|---|---|---|---|---|
| Wastewater aeration [ 150 ] | Dissolved oxygen, nitrite-to-ammonium ratio, influent load | Aeration setpoint (near-continuous) | Blower energy plus penalty on effluent limit violation | Calibrated biokinetic process model predicting the effect of aeration on nitrogen species |
| Grinding feed-rate cycle [ 292 ] | Stock removal, wheel wear, thermal and force state | Feed rate per removal increment (discretized) | Cycle time, with defect-inducing conditions entering as constraints | Virtual sensing of burn, chatter, and dimensional error along candidate cycles |
| Tire retreading and rotation [ 198 ] | Per-tire tread depth, casing condition, position in fleet | Retread, rotate, or retain (combinatorial) | Negative retread capability subject to safety and operational constraints | Learned degradation model driven by live pressure, load, and temperature data |
| UAV health-aware maneuvering [ 195 ] | Structural health state and predicted wing-strain quantities of interest | 2g or 3g maneuver | Structural-health reward combined with a reward favoring faster maneuvers | Probabilistic graphical Digital Twin that assimilates strain measurements and models the evolution of structural health under candidate maneuvers |
| Approach | Decision generation | Required Digital Twin representation | Decision horizon | Treatment of uncertainty | Online operation |
|---|---|---|---|---|---|
| Online parametric optimization | Numerically optimizes decision variables for the current operating condition | Predictive model relating candidate decisions to system responses and objectives | Single decision or application-dependent horizon | Typically uses nominal predictions without explicit uncertainty representation | Repeated numerical optimization as new system states become available |
| Nominal MPC | Optimizes a sequence of future control actions and applies the first action | Multi-step predictive dynamical model with state feedback | Finite receding horizon | Uses nominal future trajectories | Prediction and constrained optimization are repeated at each control step |
| Dynamic programming | Uses Bellman recursion to construct value functions and sequential decision policies | State-transition model, which may be deterministic or probabilistic | Finite or infinite horizon | Can explicitly account for probabilistic state transitions through expected future cost or reward | Policy synthesis can be computationally intensive, while execution can use a precomputed state-to-action policy |
| Robust MPC | Optimizes a receding-horizon control sequence that remains feasible under prescribed disturbances | Predictive dynamical model and bounded uncertainty set | Finite receding horizon | Represents uncertainty using bounded sets, tubes, or worst-case disturbances | Repeated robust optimization or constraint tightening is performed online |
| Stochastic MPC | Optimizes expected performance subject to probabilistic constraints | Predictive dynamical model and probability distributions of disturbances or future states | Finite receding horizon | Represents uncertainty probabilistically through expected objectives and chance constraints | Repeated stochastic or deterministically reformulated optimization is performed online |
| Approximate optimizer | Evaluates a precomputed or learned mapping from the current state to an action | Offline optimizer, predictive model, or simulated trajectories used to construct the policy | Typically reproduces a finite-horizon or static optimization policy | Usually inherits uncertainty treatment from the optimization problem or training data rather than quantifying it explicitly | Direct evaluation of the policy avoids repeated numerical optimization |
| Family | Learning basis/data requirement | Representation | Strengths for the decision-making | Key limitations |
|---|---|---|---|---|
| Parametric system identification, ARX/NARX [ 204 , 340 ] | Time-series simulation or sensor data | Explicit states or autoregressive histories | Interpretable dynamics, natural feedback updates, and direct integration with MPC and filtering | Limited flexibility for highly nonlinear or high-dimensional dynamics |
| Gaussian processes [ 330 , 36 ] | Typically modest paired datasets | Nonparametric input–output or state-transition mapping | Data-efficient learning with uncertainty estimates for risk-aware decisions | Computational scaling with dataset size and input dimension |
| Sparse identification, SINDy [ 190 ] | Measurements and a candidate function library | Sparse explicit dynamical equations | Interpretable dynamics with compact models for prediction and diagnosis | Sensitive to noise, state observability, and library specification |
| Latent dynamics models, autoencoders [ 379 , 423 , 141 , 260 ] | High-dimensional simulation or sensor data | Reduced coordinates or learned latent states | Reduces prediction and optimization dimension while retaining dominant system information | Reduced states may omit decision-relevant information or exhibit unstable latent dynamics |
| Neural network (NN) dynamical models, NARX encoders, recurrent neural network and time-series models [ 412 , 98 , 73 , 260 ] | Moderate to large time-series datasets | Explicit, history-dependent, or latent states | Expressive nonlinear dynamics and flexible multi-step prediction | Data-intensive, limited extrapolation, and weak rollout or stability guarantees |
| DMD and Koopman operator models [ 214 , 258 , 72 ] | Simulation or sensor trajectories | Linear dynamics in reduced or lifted coordinates | Can formulate downstream optimization as convex optimization and enable efficient MPC | Performance depends strongly on the selected observables or embedding |
| Model family | Typical inputs | Generated outputs | Suitable Digital Twin uses | Key limitations |
|---|---|---|---|---|
| Generative adversarial networks (GANs) [ 69 , 370 ] | System states, control actions, parameters, sensor measurements, and a latent sample | Images, spatial fields, fault signatures, or short trajectory segments generated in one forward pass | Fast scenario generation, virtual sensing, fault-data synthesis, and repeated evaluation within real-time monitoring or decision loops | Mode collapse may omit rare but decision-relevant operating regimes |
| Variational autoencoders (VAEs) [ 77 ] | High-dimensional, noisy, or multi-modal measurements, including sensor streams, images, and spatial fields | Reconstructed measurements, completed states, latent-state estimations | State estimation, missing-data completion, anomaly detection, and low-dimensional probabilistic state representation | Generated responses may be over-smoothed and contain less detail than GAN outputs |
| Autoregressive Transformers [ 319 , 18 ] | Sequences of past states, control actions, events, and multi-modal observations | Future states or trajectories generated sequentially over a variable prediction horizon | Long-horizon forecasting, traffic and demand prediction, equipment degradation tracking, and multi-modal sequence modeling | Inference cost increases with horizon length, and early prediction errors may propagate through the sequence |
| Diffusion models [ 153 ] | System states, controls, parameters, boundary conditions, sparse observations, and an initial noise sample | Complete trajectories, deformation or flow fields, reconstructed states, and rare or counterfactual scenarios | Uncertainty-aware planning, state reconstruction, safety testing, and generation of globally coherent multiple outcomes | Iterative denoising creates high inference cost and may preclude high-frequency control without accelerated sampling |
| Physics-informed normalizing flows [ 135 ] | Spatial coordinates, latent samples, and sparse field measurements; governing equations and boundary conditions incorporated during training | Samples of stochastic solution, coefficient, or forcing fields, including non-Gaussian distributions | Probabilistic field reconstruction, virtual sensing, inverse inference, and uncertainty propagation for decision support | Invertibility constrains architecture design; reliability depends on the validity of the encoded physics and sufficiently informative measurement snapshots |
| Type of Epistemic Model Uncertainty | Representation | Corresponding Online Adaptation Methods | Strengths | Weaknesses |
|---|---|---|---|---|
| Model parameter uncertainty | Parametric physics-based or system identification models, | Recursive parameter estimation, Bayesian calibration, Kalman filtering, and moving horizon estimation | Low-dimensional, computationally efficient, and physically interpretable updates | Limited by the predefined model structure and requires sufficiently identifiable parameters |
| Model-form uncertainty | Nominal physics-based or hybrid models with an explicit discrepancy, | Recursive bias or discrepancy correction using Gaussian processes, NNs, RNNs, or other residual models | Corrects unmodeled dynamics while retaining the knowledge and stability of the nominal model | Requires a sufficiently reliable nominal model, and parameter and discrepancy effects may be difficult to distinguish |
| Approximation uncertainty | Data-driven surrogate models, | Model fine-tuning, selective or head fine-tuning, and parameter-efficient adaptation | Provides greater plasticity for adapting learned dynamics and internal representations | Higher computational and data requirements, with risks of unstable updates and catastrophic forgetting |
Digital Twins
Digital twins are virtual models synchronized with real systems to support monitoring, prediction, simulation, and decision-making.
See all 274 papers