Pith. sign in

REVIEW 4 major objections 5 minor 106 references

A self-adaptive digital twin can stay trustworthy under concept drift by coupling Fisher-score drift detection, low-rank fine-tuning of fewer than 1% of parameters, and statistical validation before every model replacement.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 15:46 UTC pith:RJGTHPYH

load-bearing objection Novel integration of drift detection, LoRA, and validation for adaptive digital twins—useful, honestly benchmarked, but the validation gate's 'statistical certification' claim is too strong. the 4 major comments →

arxiv 2607.18164 v1 pith:RJGTHPYH submitted 2026-07-20 cs.LG cs.AImath.STstat.TH

A Continual Validation, Updating, and Decision-Making Framework for Self-Adaptive Digital Twins via Robust Model Predictive Control: A Case Study in Additive Manufacturing

Yi-Ping Chen , Ying-Kuan Tsai , Vispi Karkaria , Seul Lee , Daniel Apley , Wei Chen This is my paper
classification cs.LG cs.AImath.STstat.TH MSC 62L1062P3068T05
keywords digital twinconcept drift detectionFisher scorelow-rank adaptationonline validationrobust model predictive controlquantile regressionadditive manufacturing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes a single pipeline for keeping neural-network digital twins faithful to a drifting physical system. It answers three questions: when to update, via a Fisher-score-based multivariate drift detector; how to update, via LoRA, which tunes fewer than 1% of model parameters; and whether the update is safe, via a Mann-Whitney U test on held-out quantile loss. The framework is demonstrated under both abrupt and incremental drift in a stochastic linear system and in an additive-manufacturing process with drift embedded in material properties. The core claim is that this integration restores both predictive accuracy and uncertainty quantification, while withholding any update that cannot statistically beat the live model.

Core claim

On its own terms, the central discovery is that the zero-mean property of Fisher score vectors — the fact that at a fitted model the average gradient of the log-likelihood is zero — can be used as an online alarm for concept drift in a control loop, even when the model is trained with quantile loss rather than a true likelihood. By monitoring only the last-layer score, smoothing with multivariate EWMA, and thresholding a Hotelling T^2 statistic, the framework detects drift with short delay. LoRA then adapts the frozen pretrained surrogate from a small streaming buffer, and the Mann-Whitney U test certifies that the candidate model's quantile loss is stochastically lower before replacement. T

What carries the argument

The load-bearing object is the Fisher score vector — the gradient of the pseudo log-likelihood with respect to the model parameters, evaluated at the fitted weights — whose expectation is zero exactly when the model is correct and nonzero after concept drift. An appended last linear layer isolates the score computation; a multivariate EWMA smoother and Hotelling T^2 statistic turn the score stream into a scalar alarm; a diagonal covariance approximation and bootstrap tail-fitting make the threshold re-estimable online. LoRA provides the low-rank update W' = W + BA, with fewer than 1% of parameters trainable, and the Mann-Whitney U test acts as a statistical gate before the idle model replace

Load-bearing premise

The framework assumes that concept drift is infrequent and non-recurring, with enough time between shifts for the fine-tuned model to be collected, adapted, and validated, and that the quantile-loss gradient behaves like a true Fisher score whose mean is zero when the model is correct.

What would settle it

Run the detector on an in-control sequence where the surrogate is deliberately misspecified, for example a Gaussian-noise model on data with heavy-tailed errors, and test whether the average score vector stays zero; a systematic nonzero mean at the fitted parameters would break the zero-mean premise. Alternatively, inject two drift events closer together than the buffer-collection plus validation window and count whether the Mann-Whitney gate repeatedly rejects updates while coverage degrades.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Neural-surrogate digital twins can be kept accurate for their whole operational life without full retraining, using only small streaming batches.
  • A candidate update that fails to show a statistically significant loss reduction is withheld, so unvalidated models never enter the safety-critical control loop.
  • The drift detector can be re-initialized online via diagonal covariance and bootstrap threshold estimation, allowing repeated adaptation cycles.
  • Adaptation restores not only median predictions but also quantile bounds, so chance constraints in robust MPC remain satisfied after drift.
  • The same three-module pipeline transfers from a low-dimensional synthetic system to a high-dimensional, nonlinear manufacturing process with drift in material properties.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If quantile-loss gradients can be treated as pseudo-Fisher scores, the detector should generalize to any differentiable surrogate with a non-likelihood training loss, not only quantile-regression time-series encoders — a testable claim the paper motivates but does not fully establish.
  • The separation-of-timescales assumption suggests the framework's weakest spot is rapid successive or recurrent drift; a stress test with overlapping drift events would quantify how much responsiveness is lost.
  • Because validated low-rank updates are modular, a natural extension is to cache regime-specific LoRA weights and re-deploy them on recurrence instead of retraining — an idea the paper names as future work.
  • The Mann-Whitney validation gate is model-agnostic and could serve as a trustworthiness layer for other online adaptation schemes, such as recursive least squares or Kalman-based model updates.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes an adaptation layer for neural-network-based Digital Twins operating under concept drift. The layer combines a Fisher-score/MEWMA drift detector (Section 3.3), LoRA-based batch fine-tuning (Section 3.4), and a one-sided Mann-Whitney U test that gates model replacement (Section 3.5), built on top of a quantile-regression TiDE surrogate and robust MPC from prior work. Two studies are presented: a 30-replication linear-system benchmark with two injected drift events, and a single-run DED additive-manufacturing simulation with gradual material-property drift. The authors claim the framework detects distributional shifts with short delays, restores predictive accuracy and uncertainty quantification, and statistically certifies updates before deployment.

Significance. The integration of drift detection, parameter-efficient continual learning, and online statistical validation inside an MPC-based Digital Twin is a useful and timely contribution. Strengths include a clear ablation design (Table 2), 30-replication comparisons in the illustrative study, quantitative detection-cost timing (Fig. 27), a thoughtful online reset mechanism for the drift detector, and explicit statements of assumptions and limitations (Assumptions 1–2; acknowledgment of single-run DED results). If the validation gate were made statistically valid for dependent streaming data and the reported metrics covered the full closed-loop cycle, this would be a practically significant step toward trustworthy self-adaptive digital twins. The current manuscript falls short of its 'statistically certified update' claim because the central validation test ignores the dependence structure of the loss streams.

major comments (4)
  1. [Section 3.5, Eq. (22), Fig. 8] The Mann-Whitney U test requires independent observations within each group and independence between groups. Here the N_V=30 validation losses are consecutive outputs of a closed-loop TiDE/MPC system: they are autocorrelated (the state evolves, multi-step predictions overlap) and paired (both models see the same inputs and ground truths). Positive autocorrelation reduces the effective sample size and changes the null distribution, so the test at α=0.2 does not control the false-replacement rate. The statement that an update is 'statistically certified' before deployment is therefore unsupported. A dependence-robust procedure (e.g., block bootstrap, paired time-series test) or a substantially weaker claim is needed.
  2. [Section 4.5, Figs. 13–17] The evaluation metrics are intentionally not applied to the model-adaptation and model-validation intervals. Given the long idle periods reported in Section 4.4 (first update completed at k=456, reset at k=1156; second at k=2010), the headline ranking of ADT over stepwise baselines is based only on post-validation segments. The paper's life-cycle trustworthiness claim requires full-window metrics, or at least a quantified report of the excluded intervals; otherwise the comparison may favor the proposed framework precisely by omitting its periods of degraded performance.
  3. [Section 3.3.1 and Section 3.3.4] The theoretical guarantee behind the detector, Eqs. (14)–(16), is derived for a true likelihood and an MLE. For quantile-trained neural networks the paper treats the quantile loss as a pseudo-log-likelihood, acknowledging that 'the quantile loss does not correspond to a true likelihood' but without proving that the expected pseudo-score is zero under the null or nonzero under drift. The bootstrap UCL controls false alarms empirically, but the 'principled, if-and-only-if' detection claim does not follow. Provide a formal argument or an empirical calibration study for the pseudo-score property under quantile loss.
  4. [Section 5.7, last paragraph of Section 5] The DED case study is a single run without error bars. Quantitative comparisons such as R²=0.9168 vs 0.9064/0.8580 for unconstrained MPC and violation rates 14% vs 18%/46% for constrained MPC cannot be statistically supported from one trajectory. The authors acknowledge the single-run limitation, but the abstract and Section 5.7 present the DED results as evidence of transferability. Report multiple runs with uncertainty quantification, or explicitly restrict the case study to a qualitative demonstration.
minor comments (5)
  1. [Section 4.1, Eq. (30)] The input tuple is written as '(x^p_l, u^p_l, u^p_l)'; this appears to be a typo and should be '(x^p_l, u^p_l, u^f_l)' (or include the exogenous d sequence if used).
  2. [Section 3.3.4] The claimed inversion cost 'O(q̃^6)' is confusing: inverting a q̃×q̃ covariance matrix is O(q̃^3). Please clarify the dimensions and the complexity statement.
  3. [Table 1] Header typo: 'Hyperparamters' should be 'Hyperparameters'.
  4. [Section 4.5 and Table 2] The notation ADT LoRA-N (no validation) and ADT LoRA-Ns (smaller batch with validation) is easy to confuse. Use distinct names (e.g., ADT LoRA-NoVal and ADT LoRA-SmallBuf) and keep them consistent across text, table, and figures.
  5. [Section 6] There is a text-encoding artifact in 'Digital TwinâĂŹs'; fix to 'Digital Twin's'.

Circularity Check

0 steps flagged

No significant circularity; the derivation chain is self-contained against external benchmarks, with only minor non-load-bearing self-citations.

full rationale

The paper's claimed results are not equivalent to their inputs by construction. The drift detector is calibrated on in-control Phase I data (score mean/covariance, bootstrap UCL) and then tested on separately simulated drift, so detection delay and false alarms are not fitted outcomes. The validation gate uses a fresh held-out batch (N_V=30 or 50) collected after fine-tuning, and rejection of the Mann-Whitney null is a statistical decision rule, not a re-labeling of the model-selection loss; the reported NRMSE/coverage/NNOIS metrics are evaluated separately. The pseudo-likelihood treatment of quantile loss is explicitly acknowledged as an approximation, not a hidden identity. Self-citations ([6], [8], [62]) provide external component frameworks or a published Technometrics detection principle; they are background architecture rather than the load-bearing adaptive-validated result, which is newly integrated and benchmarked here. The Mann-Whitney dependence/autocorrelation and alpha=0.2 concerns are correctness/validity issues, not circularity. Section 6 also candidly states limitations (responsiveness, no formal MPC stability guarantees), further indicating the claims are not definitionally forced.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central claim rests on a handful of domain assumptions about drift and on a pseudo-likelihood reinterpretation of quantile loss. No new physical entities or forces are introduced, but the framework's guarantees are weaker if these assumptions fail. The manually chosen hyperparameters (λ, α, buffer sizes, validation significance) are not derived from first principles and affect performance.

free parameters (7)
  • MEWMA smoothing parameter λ = 0.05 (illustrative), 0.02 (DED)
    Controls the effective memory of the multivariate EWMA statistic; chosen by hand, directly affects detection delay and false-alarm behavior.
  • Detection threshold significance α = 1e-5 (illustrative), 1e-4 (DED)
    Used to set the upper control limit for the Hotelling T² statistic; arbitrary but conventional, not derived from data.
  • Mann-Whitney U test significance level = 0.2
    Unusually high significance level for validation; the paper says it is due to limited validation set size, but this weakens the 'statistically certifies' claim.
  • LoRA rank r = 1
    Rank of low-rank adaptation; chosen as a design knob for the stability-plasticity trade-off, not optimized exhaustively.
  • Fine-tuning buffer size N_B = 200 (illustrative), 300 (DED)
    Number of streaming samples used for each batch-wise fine-tuning round; affects generalization and adaptation responsiveness.
  • Validation buffer size N_V = 30 (illustrative), 50 (DED)
    Number of samples used for the online Mann-Whitney U validation; directly limits statistical power.
  • Fine-tuning learning rate, regularization, epochs = 1e-3, 2e-4, 100
    Chosen from preliminary experiments to ensure stable convergence; not exhaustively optimized and acknowledged as case-dependent.
axioms (6)
  • domain assumption Assumption 1: The physical system encounters only abrupt drift and incremental drift, which are non-recurring and infrequent.
    Adopted in Section 3 to motivate the batch-wise, event-triggered design; if drift is recurrent or frequent, the framework may lose responsiveness.
  • domain assumption Assumption 2: After model adaptation, the system does not experience immediate subsequent drastic drift; drift evolves on a timescale allowing the updated model to remain valid.
    Ensures that the updated model and reinitialized detector operate long enough to be useful; violated in rapidly changing environments.
  • ad hoc to paper Quantile loss can be treated as a pseudo-log-likelihood, and its gradient behaves like a Fisher score vector.
    Invoked in Section 3.3.4 to apply score-based drift detection to networks trained with quantile loss; the zero-mean property is exact for true MLEs but only approximate here.
  • domain assumption Monitoring the score vector of the last layer is sufficient to detect concept drift in the conditional distribution p(y|x).
    Used in Section 3.3.3, following Zhang and Apley [62], to reduce computational cost; may miss changes in learned feature representations.
  • domain assumption The validation set collected after adaptation is representative of the future streaming distribution.
    The Mann-Whitney U test in Section 3.5 compares losses on a post-adaptation batch; if covariate shift continues, validation may not reflect future performance.
  • domain assumption GAMMA FEA simulation is a faithful representation of the physical DED process for evaluating the framework.
    The engineering case study relies entirely on this in-house simulation; the authors acknowledge it ignores phase change and fluid-flow effects, requiring calibration.

pith-pipeline@v1.3.0-alltime-deepseek · 37262 in / 7200 out tokens · 68804 ms · 2026-08-01T15:46:22.681499+00:00 · methodology

0 comments
read the original abstract

Digital Twins rely on surrogate models to mirror physical systems in real time, yet these models can degrade as operating conditions evolve, a phenomenon known as concept drift. Maintaining surrogate fidelity under drift, particularly when models must also capture aleatoric uncertainty, remains an open challenge. Existing adaptive frameworks lack principled mechanisms for detecting when updates are needed, for efficiently adapting models from limited streaming data, and for certifying that updates genuinely improve predictive performance. Here we present an adaptive Digital Twin framework that integrates a Fisher score--based multivariate drift detector, Low-Rank Adaptation (LoRA) for parameter-efficient continual learning, and a Mann--Whitney $U$ test for online statistical validation. The framework monitors surrogate-model confidence via Fisher score vectors, triggers targeted fine-tuning of fewer than 1% of model parameters upon drift detection, and statistically certifies predictive improvement before deploying the updated surrogate. Applied to a stochastic linear system and a directed energy deposition additive manufacturing process as case studies, the framework successfully detects distributional shifts with short delays and restores both predictive accuracy and uncertainty quantification under abrupt and incremental drift. These results establish a statistically rigorous and computationally tractable pathway for sustaining the trustworthiness of neural-network--based Digital Twins throughout their operational life cycle.

Figures

Figures reproduced from arXiv: 2607.18164 by Daniel Apley, Seul Lee, Vispi Karkaria, Wei Chen, Ying-Kuan Tsai, Yi-Ping Chen.

Figure 1
Figure 1. Figure 1: The proposed framework for the Adaptive Digital Twin. The gray box and the blue box on the left-hand side [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of MPC and robust MPC. (a) Illustrates MPC at time [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The network structure of Time Series Dense Encoder (TiDE), modified from [ [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The data structure of TiDE 2.3 Quantile Regression In engineering applications, a predictive model that returns only a single point estimate of future system states is often insufficient for robust decision-making, as it gives no indication of how reliable that estimate is under disturbances. Quantile regression [89–91] addresses this limitation by predicting conditional quantiles rather than the mean, yie… view at source ↗
Figure 5
Figure 5. Figure 5: Timeline of the proposed framework [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of the score-based drift detection principle. Following the optimality condition, the MLE [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of fine-tuning methods. Low-rank adaptation (LoRA) [69], a widely used PEFT method for large language models, constrains task-specific adaptation to a low-rank subspace of the pretrained weights, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Flowchart of applying Mann-Whitney U test for model validation. [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Trajectory of RMPC in the in-control condition for setting up the drift detector in Phase I [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Hotelling 𝑇 2 statistics under in-control conditions for upper-quantile estimation. (a) The 𝑇 2 trajectory of the streaming data. Under stationary conditions, the 𝑇 2 statistic fluctuates around a stable level and remains below the threshold, indicating the absence of false alarms. (b) Empirical distribution of 𝑇 2 in Phase I, well approximated by the fitted scaled noncentral 𝜒 2 distribution. 4.4 Phase I… view at source ↗
Figure 11
Figure 11. Figure 11: Trajectories of the adaptive Digital Twin in Phase II. The figure shows the trajectories of [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Loss curves during fine-tuning, histograms of the quantile loss at the validation stage, and the estimated [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Distribution of the normalized root mean square error (NRMSE) across 30 replications for each benchmark [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Empirical coverage of the 90% prediction interval across 30 replications for each method, for states 𝑥1 (top) and 𝑥2 (bottom) under (a) 𝑃1 (𝑦 | 𝑥) and (b) 𝑃2 (𝑦 | 𝑥). Values closer to the nominal level of 0.9 indicate better-estimated quantiles, while lower values reflect under-coverage. 5.1 System Overview and Scope The Digital Twin is designed for controlling melt pool dynamics, the most critical and co… view at source ↗
Figure 15
Figure 15. Figure 15: Distribution of the normalized negatively oriented interval score (NNOIS) across 30 replications for each [PITH_FULL_IMAGE:figures/full_fig_p027_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Distribution of the quantile loss 𝐿𝑄 across 30 replications for each method, for states 𝑥1 (top) and 𝑥2 (bottom) under (a) 𝑃1 (𝑦 | 𝑥) and (b) 𝑃2 (𝑦 | 𝑥). Lower values indicate closer agreement between the predicted quantiles and the realized responses. 5.2 Physical System: GAMMA Simulation of DED Process The physical DED process is represented by GAMMA, an in-house GPU-accelerated explicit FEA solver [102… view at source ↗
Figure 17
Figure 17. Figure 17: Average performance ranking of all benchmark methods across the four evaluation metrics (NNOIS, NRMSE, [PITH_FULL_IMAGE:figures/full_fig_p028_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Single-track square [PITH_FULL_IMAGE:figures/full_fig_p028_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Trajectory of RMPC in the in-control condition for setting up the drift detector in Phase I. The RMPC framework [PITH_FULL_IMAGE:figures/full_fig_p030_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Hotelling 𝑇 2 statistics under in-control conditions for upper-quantile estimation. (a) The 𝑇 2 trajectory of the streaming data. (b) Empirical distribution of 𝑇 2 in Phase I, well approximated by the fitted scaled noncentral 𝜒 2 distribution. during which the weight of 316L is gradually reduced by 1.5% and replaced by AISI 1018 over 40 steps. This reduction continues until time step 2200, reaching a tota… view at source ↗
Figure 21
Figure 21. Figure 21: Temperature-dependent material properties for stainless steel 316L, AISI 1018, and their mixture. (a) [PITH_FULL_IMAGE:figures/full_fig_p031_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Resulting trajectories of melt pool temperature, melt pool depth, and laser power input in the DED printing [PITH_FULL_IMAGE:figures/full_fig_p032_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: TiDE-predicted median and quantiles under different operating stages. (a) shows the predictions under nominal, [PITH_FULL_IMAGE:figures/full_fig_p033_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Loss curves during the fine-tuning stages, histograms of the quantile loss at the validation stage, and the [PITH_FULL_IMAGE:figures/full_fig_p033_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Resulting trajectories of melt pool temperature, melt pool depth, and laser power under constrained robust MPC, [PITH_FULL_IMAGE:figures/full_fig_p035_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Fine-tuning loss curves, validation-stage quantile-loss histograms (live vs. idle model), and the estimated [PITH_FULL_IMAGE:figures/full_fig_p036_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Histogram of the CPU time of the score-based drift detector [PITH_FULL_IMAGE:figures/full_fig_p036_27.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

106 extracted references · 7 linked inside Pith

  1. [1]

    Digital twin: Mitigating unpredictable, undesirable emergent behavior in complex systems

    Grieves, M., and Vickers, J., 2016. “Digital twin: Mitigating unpredictable, undesirable emergent behavior in complex systems”. InTransdisciplinary perspectives on complex systems: New findings and approaches. Springer, pp. 85–113

    2016

  2. [2]

    National Academies of Sciences, Engineering, and Medicine and others, 2023.Foundational research gaps and future directions for digital twins

    2023

  3. [3]

    Digital twins for the designs of systems: a perspective

    van Beek, A., Nevile Karkaria, V., and Chen, W., 2023. “Digital twins for the designs of systems: a perspective”. Structural and Multidisciplinary Optimization,66(3), p. 49

    2023

  4. [4]

    Anoptimization-centricreviewforintegratingartificial intelligence and digital twin technologies in manufacturing

    Karkaria,V.,Tsai,Y.-K.,Chen,Y.-P.,andChen,W.,2025. “Anoptimization-centricreviewforintegratingartificial intelligence and digital twin technologies in manufacturing”.Engineering Optimization, pp. 1–47

    2025

  5. [5]

    Decision support within digital twins in manufacturing ecosystems: A review

    Gálvez del Postigo Gallego, L., Lazarova-Molnar, S., del Real Torres, A., and Acevedo Galicia, L. E., 2025. “Decision support within digital twins in manufacturing ecosystems: A review”.IET Collaborative Intelligent Manufacturing,7(1), p. e70041

    2025

  6. [6]

    Uncertainty-aware digital twins: Robust model predictive control using time-series deep quantile learning

    Chen, Y.-P., Tsai, Y.-K., Karkaria, V., and Chen, W., 2026. “Uncertainty-aware digital twins: Robust model predictive control using time-series deep quantile learning”.Journal of Mechanical Design,148(2), p. 021702

    2026

  7. [7]

    Amodulardigitaltwin architecture for offshore safety and decision-making: A north sea platform case study

    Hamzah,N.,Mutallimov,T.,Panaitescu,C.-T.,Grant,M.,Cristofoli,G.,andTan,H.,2026. “Amodulardigitaltwin architecture for offshore safety and decision-making: A north sea platform case study”.Reliability Engineering & System Safety, p. 112428

    2026

  8. [8]

    Real-time decision-making for digital twin in additive manufacturing with model predictive control using time-series deep neural networks

    Chen, Y.-P., Karkaria, V., Tsai, Y.-K., Rolark, F., Quispe, D., Gao, R. X., Cao, J., and Chen, W., 2025. “Real-time decision-making for digital twin in additive manufacturing with model predictive control using time-series deep neural networks”.Journal of Manufacturing Systems,80, pp. 412–424

    2025

  9. [9]

    The emergence of cognitive digital twin: vision, challenges and opportunities

    Zheng, X., Lu, J., and Kiritsis, D., 2022. “The emergence of cognitive digital twin: vision, challenges and opportunities”.International Journal of Production Research,60(24), pp. 7610–7632

    2022

  10. [10]

    Maintenance planning using a digital twin: principles and case studies

    Dwight, R., Li, W., van Rooij, F., and Scarf, P., 2025. “Maintenance planning using a digital twin: principles and case studies”.Reliability Engineering & System Safety, p. 111496

    2025

  11. [11]

    Ai-driven digital twin framework for predictive maintenance: Dynamic decision making and adaptive model updates

    Karkaria, V. N., Chen, Y.-P., Tsai, Y.-K., Erdogan, G., Radulescu, R., and Chen, W., 2025. “Ai-driven digital twin framework for predictive maintenance: Dynamic decision making and adaptive model updates”. In International DesignEngineeringTechnicalConferencesandComputersandInformationinEngineeringConference,Vol.89213, American Society of Mechanical Engin...

    2025

  12. [12]

    Adaptive digital twin of sheet metal forming via proper orthogonal decomposition-based koopman operator with model predictive control

    Chen,Y.-P.,Suarez,D.,Tsai,Y.-K.,Karkaria,V.,Hu,G.,Chen,Z.,Guo,P.,Cao,J.,andChen,W.,2025. “Adaptive digital twin of sheet metal forming via proper orthogonal decomposition-based koopman operator with model predictive control”.arXiv preprint arXiv:2511.10852

    Pith/arXiv arXiv 2025

  13. [13]

    A comprehensive review of digital twinâĂŤpart 1: modeling and twinning enabling technologies

    Thelen, A., Zhang, X., Fink, O., Lu, Y., Ghosh, S., Youn, B. D., Todd, M. D., Mahadevan, S., Hu, C., and Hu, Z., 2022. “A comprehensive review of digital twinâĂŤpart 1: modeling and twinning enabling technologies”. Structural and Multidisciplinary Optimization,65(12), p. 354

    2022

  14. [14]

    Towards a digital twin framework in additive manufacturing: Machine learning and bayesian optimization for time series process optimization

    Karkaria,V.,Goeckner,A.,Zha,R.,Chen,J.,Zhang,J.,Zhu,Q.,Cao,J.,Gao,R.X.,andChen,W.,2024. “Towards a digital twin framework in additive manufacturing: Machine learning and bayesian optimization for time series process optimization”.Journal of Manufacturing Systems

    2024

  15. [15]

    Learning dynamical systems from data: An introduction to physics-guided deep learning

    Yu, R., and Wang, R., 2024. “Learning dynamical systems from data: An introduction to physics-guided deep learning”.Proceedings of the National Academy of Sciences,121(27), p. e2311808121. 38

    2024

  16. [16]

    2026 roadmap on artificial intelligence and machine learning for smart manufacturing

    Lee, J., Su, H., Macchi, M., Polenghi, A., Wu, W., Zhao, Z., Huang, G. Q., Allgood, K., Jain, D., Gieger, B., Pandhare, V., Bhattacharjee, S., Mohril, R., Kong, L., Wang, Q., Tang, X., Kim, S., Park, C. H., Youn, B. D., Goh, G. D., Huang, X., Yeong, W. Y., Shin, Y. C., Zhang, H., Wang, Z., Tao, F., Srai, J. S., Gupta, S. K., Joung, B. G., John, A., Suther...

    2026

  17. [17]

    Digitaltwin-enabledmultigenerationcontrolco-design with deep reinforcement learning

    Tsai,Y.-K.,Karkaria,V.,Chen,Y.-P.,andChen,W.,2026. “Digitaltwin-enabledmultigenerationcontrolco-design with deep reinforcement learning”.Journal of Mechanical Design,148(7), 06, p. 071705

    2026

  18. [18]

    Reinforcement learning-based control co-design of digital twin-enabled full-vehicle active suspension systems

    Tsai, Y.-K., Chen, Y.-P., Karkaria, V., and Chen, W., 2026. “Reinforcement learning-based control co-design of digital twin-enabled full-vehicle active suspension systems”.Structural and Multidisciplinary Optimization

    2026

  19. [19]

    Methods for enabling real-time analysis in digital twins: A literature review

    Es-haghi, M. S., Anitescu, C., and Rabczuk, T., 2024. “Methods for enabling real-time analysis in digital twins: A literature review”.Computers & Structures,297, p. 107342

    2024

  20. [20]

    Model predictive control: theory, computation, and design

    Rawlings, J. B., Mayne, D. Q., and Diehl, M. M., 2020. “Model predictive control: theory, computation, and design”.(No Title)

    2020

  21. [21]

    Learning-based model predictive control: Towardsafelearningincontrol

    Hewing, L., Wabersich, K. P., Menner, M., and Zeilinger, M. N., 2020. “Learning-based model predictive control: Towardsafelearningincontrol”.AnnualReviewofControl,Robotics,andAutonomousSystems,3(1),pp.269–296

    2020

  22. [22]

    Robustcontrolco-designusingtube-basedmodelpredictivecontrol

    Tsai,Y.-K.,andMalak,R.J.,2023. “Robustcontrolco-designusingtube-basedmodelpredictivecontrol”. In2023 American Control Conference (ACC), IEEE, pp. 769–775

    2023

  23. [23]

    Control co-design with performance-robustness trade-off using tube-based stochastic model predictive control

    Tsai, Y.-K., and Malak, R., 2025. “Control co-design with performance-robustness trade-off using tube-based stochastic model predictive control”.ASME Letters in Dynamic Systems and Control

    2025

  24. [24]

    Parametric multi-objective optimization for simultaneous and nested control co-design formulations with tube-based model predictive controllers

    Tsai, Y.-K., and Malak, R. J., 2026. “Parametric multi-objective optimization for simultaneous and nested control co-design formulations with tube-based model predictive controllers”.Journal of Mechanical Design, pp. 1–18

    2026

  25. [25]

    Simultaneous multistep transformer architecture for model predictive control

    Park, J., Babaei, M. R., Munoz, S. A., Venkat, A. N., and Hedengren, J. D., 2023. “Simultaneous multistep transformer architecture for model predictive control”.Computers & Chemical Engineering,178, p. 108396

    2023

  26. [26]

    Lstm-mpc: A deep learning based predictive control method for multimode process control

    Huang, K., Wei, K., Li, F., Yang, C., and Gui, W., 2022. “Lstm-mpc: A deep learning based predictive control method for multimode process control”.IEEE Transactions on Industrial Electronics,70(11), pp. 11544–11554

    2022

  27. [27]

    Model predictive control when utilizing lstm as dynamic models

    Jung, M., da Costa Mendes, P. R., Önnheim, M., and Gustavsson, E., 2023. “Model predictive control when utilizing lstm as dynamic models”.Engineering Applications of Artificial Intelligence,123, p. 106226

    2023

  28. [28]

    Process structure-based recurrent neural network modeling for model predictive control of nonlinear processes

    Wu, Z., Rincon, D., and Christofides, P. D., 2020. “Process structure-based recurrent neural network modeling for model predictive control of nonlinear processes”.Journal of Process Control,89, pp. 74–84

    2020

  29. [29]

    Anewadaptiveexplicitnonlinearmodelpredictivecontroldesignforanonlinear mimo system: An application to twin rotor mimo system

    Dutta,L.,andDas,D.K.,2021. “Anewadaptiveexplicitnonlinearmodelpredictivecontroldesignforanonlinear mimo system: An application to twin rotor mimo system”.International Journal of Control, Automation and Systems,19(7), pp. 2406–2419

    2021

  30. [30]

    Targeted digital twin via flow map learning and its application to fluid dynamics

    Chen, Q., Xu, Z., Zhang, J., and Xiu, D., 2025. “Targeted digital twin via flow map learning and its application to fluid dynamics”.arXiv preprint arXiv:2510.07549

    arXiv 2025

  31. [31]

    Unknown parameter excitation and estimation for complex systems with dynamic performances

    Chen, Y.-P., and Chan, K.-Y., 2021. “Unknown parameter excitation and estimation for complex systems with dynamic performances”.Journal of Mechanical Design,143(9), p. 091704

    2021

  32. [32]

    Cautious model predictive control using gaussian process regression

    Hewing, L., Kabzan, J., and Zeilinger, M. N., 2019. “Cautious model predictive control using gaussian process regression”.IEEE Transactions on Control Systems Technology,28(6), pp. 2736–2743. 39

    2019

  33. [33]

    Bias-learning-based model predictive controller design for reliable path tracking of autonomous vehicles under model and environmental uncertainty

    Ren, L., and Xi, Z., 2022. “Bias-learning-based model predictive controller design for reliable path tracking of autonomous vehicles under model and environmental uncertainty”.Journal of Mechanical Design,144(9), p. 091706

    2022

  34. [34]

    Self-adaptive fault diagnosis for unseen working conditions based on digital twins and domain generalization

    Azari, M. S., Santini, S., Edrisi, F., and Flammini, F., 2025. “Self-adaptive fault diagnosis for unseen working conditions based on digital twins and domain generalization”.Reliability Engineering & System Safety,254, p. 110560

    2025

  35. [35]

    Investigation on the impact of nongeometric uncertainty in dynamic perfor- mance of serial and parallel robot manipulators

    Tsai, Y.-K., and Chan, K.-Y., 2019. “Investigation on the impact of nongeometric uncertainty in dynamic perfor- mance of serial and parallel robot manipulators”.Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science,233(10), pp. 3487–3511

    2019

  36. [36]

    A novel digital twin approach based on deep multimodal information fusion for aero-engine fault diagnosis

    Huang, Y., Tao, J., Sun, G., Wu, T., Yu, L., and Zhao, X., 2023. “A novel digital twin approach based on deep multimodal information fusion for aero-engine fault diagnosis”.Energy,270, p. 126894

    2023

  37. [37]

    An accelerated dual gradient-projection algorithm for embedded linear model predictive control

    Patrinos, P., and Bemporad, A., 2013. “An accelerated dual gradient-projection algorithm for embedded linear model predictive control”.IEEE Transactions on Automatic Control,59(1), pp. 18–33

    2013

  38. [38]

    Mpc-based dual control with online experiment design

    Heirung, T. A. N., Foss, B., and Ydstie, B. E., 2015. “Mpc-based dual control with online experiment design”. Journal of Process Control,32, pp. 64–76

    2015

  39. [39]

    Adaptiveoptimalcontroloffaultynonlinear dc microgrids with constant power loads: Dual-extended kalman filter approach

    Vafamand,N.,Arefi,M.M.,Shafie-Khah,M.,andCatalão,J.P.,2022. “Adaptiveoptimalcontroloffaultynonlinear dc microgrids with constant power loads: Dual-extended kalman filter approach”.IEEE Transactions on Industry Applications,59(1), pp. 513–522

    2022

  40. [40]

    Koopman operator-based model predictive control with recursive online update

    Calderón, H. M., Schulz, E., Oehlschlägel, T., and Werner, H., 2021. “Koopman operator-based model predictive control with recursive online update”. In 2021 European Control Conference (ECC), IEEE, pp. 1543–1549

    2021

  41. [41]

    A sample path-based method for the digital twin prediction update synchronizationproblemofunreliableproductionlines

    Gou, Y.-X., Matta, A., and Tan, B., 2024. “A sample path-based method for the digital twin prediction update synchronizationproblemofunreliableproductionlines”.In2024IEEE20thInternationalConferenceonAutomation Science and Engineering (CASE), IEEE, pp. 2969–2974

    2024

  42. [42]

    Understanding semantic drift in model driven digital twins

    Abbasi, F., Brimont, P., Pruski, C., and Sottet, J.-S., 2024. “Understanding semantic drift in model driven digital twins”. In Proceedings of the ACM/IEEE 27th International Conference on Model Driven Engineering Languages and Systems, pp. 419–430

    2024

  43. [43]

    Springer Nature

    Rabczuk, T., Anitescu, C., Goswami, S., Zhuang, X., and Wang, Y., 2026.Scientific Machine Learning with Engineering Applications. Springer Nature

    2026

  44. [44]

    Adaptive neural network model predictive control

    Hedjar, R., 2013. “Adaptive neural network model predictive control”.International Journal of Innovative Computing, Information and Control,9(3), pp. 1245–1257

    2013

  45. [45]

    Nonlinear model identification and adaptive model predictive control using neural networks

    Akpan, V. A., and Hassapis, G. D., 2011. “Nonlinear model identification and adaptive model predictive control using neural networks”.ISA transactions,50(2), pp. 177–194

    2011

  46. [46]

    Real-time adaptive machine-learning-based predictive control of nonlinear processes

    Wu, Z., Rincon, D., and Christofides, P. D., 2019. “Real-time adaptive machine-learning-based predictive control of nonlinear processes”.Industrial & Engineering Chemistry Research,59(6), pp. 2275–2290

    2019

  47. [47]

    An online learning method for microgrid energy management control

    Casagrande, V., Ferianc, M., Rodrigues, M., and Boem, F., 2023. “An online learning method for microgrid energy management control”. In 2023 31st Mediterranean Conference on Control and Automation (MED), IEEE, pp. 263–268

    2023

  48. [48]

    Adaptivemachinelearningwithphysics-basedsimulationsfor mean time to failure prediction of engineering systems

    Wu,H.,Xu,Y.,Liu,Z.,Li,Y.,andWang,P.,2023. “Adaptivemachinelearningwithphysics-basedsimulationsfor mean time to failure prediction of engineering systems”.Reliability Engineering & System Safety,240, p. 109553

    2023

  49. [49]

    Digitaltwin-basedanomalydetectionunderconceptdrift: Acom- parison between iterative batch learning and incremental learning approaches

    Abdoune,F.,Nouiri,M.,andCardin,O.,2025. “Digitaltwin-basedanomalydetectionunderconceptdrift: Acom- parison between iterative batch learning and incremental learning approaches”.Expert Systems with Applications, p. 130062. 40

    2025

  50. [50]

    Deep model predictive flow control with limited sensor data and online learning

    Bieker, K., Peitz, S., Brunton, S. L., Kutz, J. N., and Dellnitz, M., 2020. “Deep model predictive flow control with limited sensor data and online learning”.Theoretical and computational fluid dynamics,34(4), pp. 577–591

    2020

  51. [51]

    Anonlinetransferlearningapproachforidentification and predictive control design with application to rcci engines

    Bao,Y.,MohammadpourVelni,J.,andShahbakhti,M.,2020.“Anonlinetransferlearningapproachforidentification and predictive control design with application to rcci engines”. In Dynamic systems and control conference, Vol. 84270, American Society of Mechanical Engineers, p. V001T21A003

    2020

  52. [52]

    Model predictive control with adaptive machine-learning-based model for building energy efficiency and comfort optimization

    Yang, S., Wan, M. P., Chen, W., Ng, B. F., and Dubey, S., 2020. “Model predictive control with adaptive machine-learning-based model for building energy efficiency and comfort optimization”.Applied Energy,271, p. 115147

    2020

  53. [53]

    Concept drift detection and adaption framework using optimized deep learning and adaptive sliding window approach

    Desale, K. S., and Shinde, S. V., 2023. “Concept drift detection and adaption framework using optimized deep learning and adaptive sliding window approach”.Expert Systems,40(9), p. e13394

    2023

  54. [54]

    A lightweight concept drift detection and adaptation framework for iot data streams

    Yang, L., and Shami, A., 2021. “A lightweight concept drift detection and adaptation framework for iot data streams”.IEEE Internet of Things Magazine,4(2), pp. 96–101

    2021

  55. [55]

    A drift detection method based on dynamic classifier selection

    Felipe, P., dos Santos Eulanda, M., and João, G., 2020. “A drift detection method based on dynamic classifier selection”.Data Mining and Knowledge Discovery,34(1), pp. 50–74

    2020

  56. [56]

    Asurveyonconceptdriftadaptation

    Gama,J.,Žliobait ˙e,I.,Bifet,A.,Pechenizkiy,M.,andBouchachia,A.,2014. “Asurveyonconceptdriftadaptation”. ACM computing surveys (CSUR),46(4), pp. 1–37

    2014

  57. [57]

    Learning under concept drift: A review

    Lu, J., Liu, A., Dong, F., Gu, F., Gama, J., and Zhang, G., 2018. “Learning under concept drift: A review”.IEEE transactions on knowledge and data engineering,31(12), pp. 2346–2363

    2018

  58. [58]

    Learningfromstreamingdatawithconceptdriftandimbalance: an overview

    Hoens,T.R.,Polikar,R.,andChawla,N.V.,2012. “Learningfromstreamingdatawithconceptdriftandimbalance: an overview”.Progress in artificial intelligence,1(1), pp. 89–101

    2012

  59. [59]

    A comparative study on concept drift detectors

    Gonçalves Jr, P. M., de Carvalho Santos, S. G., Barros, R. S., and Vieira, D. C., 2014. “A comparative study on concept drift detectors”.Expert Systems with Applications,41(18), pp. 8144–8156

    2014

  60. [60]

    An ewma for monitoring a process standard deviation

    Crowder, S. V., and Hamilton, M. D., 1992. “An ewma for monitoring a process standard deviation”.Journal of Quality Technology,24(1), pp. 12–21

    1992

  61. [61]

    Quality control of surgical and interventional procedures: a review of the cusum

    Biau, D. J., Resche-Rigon, M., Godiris-Petit, G., Nizard, R. S., and Porcher, R., 2007. “Quality control of surgical and interventional procedures: a review of the cusum”.BMJ Quality & Safety,16(3), pp. 203–207

    2007

  62. [62]

    Concept drift monitoring and diagnostics of supervised learning models via score vectors

    Zhang, K., Bui, A. T., and Apley, D. W., 2023. “Concept drift monitoring and diagnostics of supervised learning models via score vectors”.Technometrics,65(2), pp. 137–149

    2023

  63. [63]

    Learning from time-changing data with adaptive windowing

    Bifet, A., and Gavalda, R., 2007. “Learning from time-changing data with adaptive windowing”. In Proceedings of the 2007 SIAM international conference on data mining, SIAM, pp. 443–448

    2007

  64. [64]

    Integrating structured biological data by kernel maximum mean discrepancy

    Borgwardt, K. M., Gretton, A., Rasch, M. J., Kriegel, H.-P., Schölkopf, B., and Smola, A. J., 2006. “Integrating structured biological data by kernel maximum mean discrepancy”.Bioinformatics,22(14), pp. e49–e57

    2006

  65. [65]

    Adaptive monitoring method for batch processes based onphasedissimilarityupdatingwithlimitedmodelingdata

    Zhao, C., Wang, F., Gao, F., Lu, N., and Jia, M., 2007. “Adaptive monitoring method for batch processes based onphasedissimilarityupdatingwithlimitedmodelingdata”.Industrial&engineeringchemistryresearch,46(14), pp. 4943–4953

    2007

  66. [66]

    Overcoming catastrophic forgetting in neural networks

    Kirkpatrick,J.,Pascanu,R.,Rabinowitz,N.,Veness,J.,Desjardins,G.,Rusu,A.A.,Milan,K.,Quan,J.,Ramalho, T., Grabska-Barwinska, A., et al., 2017. “Overcoming catastrophic forgetting in neural networks”.Proceedings of the national academy of sciences,114(13), pp. 3521–3526

    2017

  67. [67]

    Catastrophic forgetting in deep learning: A comprehensive taxonomy

    Aleixo, E. L., Colonna, J. G., Cristo, M., and Fernandes, E., 2023. “Catastrophic forgetting in deep learning: A comprehensive taxonomy”.arXiv preprint arXiv:2312.10549. 41

    Pith/arXiv arXiv 2023

  68. [68]

    Parameter-efficient fine-tuning for large models: A comprehensive survey

    Han, Z., Gao, C., Liu, J., Zhang, J., and Zhang, S. Q., 2024. “Parameter-efficient fine-tuning for large models: A comprehensive survey”.arXiv preprint arXiv:2403.14608

    Pith/arXiv arXiv 2024

  69. [69]

    Lora: Low-rank adaptation of large language models

    Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al., 2022. “Lora: Low-rank adaptation of large language models.”.Iclr,1(2), p. 3

    2022

  70. [70]

    Adaptive parameterized model predictive control based on reinforcement learning: A synthesis framework

    Sun, D., Jamshidnejad, A., and De Schutter, B., 2024. “Adaptive parameterized model predictive control based on reinforcement learning: A synthesis framework”.Engineering Applications of Artificial Intelligence,136, p. 109009

    2024

  71. [71]

    Output-feedback robust control of uncertain systems via online data-driven learning

    Na, J., Zhao, J., Gao, G., and Li, Z., 2020. “Output-feedback robust control of uncertain systems via online data-driven learning”.IEEE Transactions on Neural Networks and Learning Systems,32(6), pp. 2650–2662

    2020

  72. [72]

    Online weight-adaptive nonlinear model predictive control

    Kostadinov, D., and Scaramuzza, D., 2020. “Online weight-adaptive nonlinear model predictive control”. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE, pp. 1180–1185

    2020

  73. [73]

    An analytical tuning of mpc control horizon using the hessian condition number

    Turki, M., Langlois, N., and Yassine, A., 2017. “An analytical tuning of mpc control horizon using the hessian condition number”. In 14th International Workshop on Advanced Control and Diagnosis

    2017

  74. [74]

    Neural network approach to continuous-time direct adaptive optimal control for partially unknown nonlinear systems

    Vrabie, D., and Lewis, F., 2009. “Neural network approach to continuous-time direct adaptive optimal control for partially unknown nonlinear systems”.Neural networks,22(3), pp. 237–246

    2009

  75. [75]

    Adaptivemodelpredictivecontrolformulti-agentsystemsusing twin delayed deep deterministic policy gradient learning

    Chaubey,P.,Markana,A.,andVyas,D.R.,2026. “Adaptivemodelpredictivecontrolformulti-agentsystemsusing twin delayed deep deterministic policy gradient learning”.Engineering Research Express

    2026

  76. [76]

    Digital twin syncing for autonomous surface vessels using reinforcement learning and nonlinear model predictive control

    Berg, H. S., Menges, D., Tengesdal, T., and Rasheed, A., 2025. “Digital twin syncing for autonomous surface vessels using reinforcement learning and nonlinear model predictive control”.Scientific Reports,15(1), p. 9344

    2025

  77. [77]

    A self-calibrating digital twin for space thermionic nuclear reactors under fault conditions

    Zhang, B., Xue, Y., Zhang, Y., and Su, K., 2025. “A self-calibrating digital twin for space thermionic nuclear reactors under fault conditions”.Reliability Engineering & System Safety, p. 111995

    2025

  78. [78]

    Dual adaptive model predictive control

    Heirung, T. A. N., Ydstie, B. E., and Foss, B., 2017. “Dual adaptive model predictive control”.Automatica,80, pp. 340–348

    2017

  79. [79]

    Learning-basedmodelpredictivecontrolonaquadrotor: Onboard implementation and experimental results

    Bouffard,P.,Aswani,A.,andTomlin,C.,2012. “Learning-basedmodelpredictivecontrolonaquadrotor: Onboard implementation and experimental results”. In 2012 IEEE International Conference on Robotics and Automation, IEEE, pp. 279–284

    2012

  80. [80]

    Real-time multivariable control of directed energy deposition via adaptive model predictive control

    van Blitterswijk, R. H., Botelho, L. A., and Khajepour, A., 2025. “Real-time multivariable control of directed energy deposition via adaptive model predictive control”.Additive Manufacturing, p. 104941

    2025

Showing first 80 references.