Equation-Free Digital Twins for Nonlinear Structural Dynamics

arXiv:2605.00950 · eess.SP, cs.LG, cs.SY, eess.SY · Submitted 2026-05-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Equation-Free Digital Twins for Nonlinear Structural Dynamics".

Dev: Here's the gist of the paper. Monitoring high-dimensional engineering structures in extreme environments is fundamentally limited not only by nonlinearity, but also by the convergence of non-stationary excitation, nonlinear structural kinematics, and complex stochastic forcing <ref:2605.00950#pg2> This paper introduces a rank-optimized Digital Twin framework based on Koopman Operator Theory, utilizing Hankel-matrix embeddings and dynamic mode decomposition to bridge the gap between high-fidelity physics and real-time edge processing <ref:2605.00950#pg2> By lifting operational data into a linear invariant subspace, the model achieves an autonomous, input-blind reconstruction of the structural state without requiring a priori mass or stiffness matrices <ref:2605.00950#pg2> Framework Overview The proposed framework is an equation-free digital twin based on Koopman Operator Theory that bridges high-fidelity physics and real-time edge processing <ref:2605.00950#pg2> It achieves this by lifting nonlinear dynamics into a higher-dimensional functional space where their evolution is linear <ref:2605.00950#pg2> The core objective is to find a finite-dimensional approximation of the Koopman operator, denoted as A, which governs the time evolution of available sensor measurements yk in R p <ref:2605.00950#pg2> This approach eliminates the need for a priori mass-stiffness matrices while maintaining physical interpretability <ref:2605.00950#pg2> How the training signal is built The training dataset was generated using OpenFAST, modeling the NREL 5MW wind turbine on the OC3 Hywind Spar platform <ref:2605.00950#pg2> The simulation setup was explicitly designed to capture the chaotic, coupled nature of the offshore environment essential for identification <ref:2605.00950#pg2> To ensure a robust training set, a randomized training ensemble was constructed by sampling from the global dataset, preventing overfitting to specific time-histories <ref:2605.00950#pg2> The state vector is constructed by fusing two distinct physical quantities: absolute acceleration and bending moment (derived from strain) <ref:2605.00950#pg2> Where the gains come from The rank-optimized manifold effectively unmixes structural resonances from deterministic 3P rotor harmonics, a task where standard subspace identification typically fails under colored noise <ref:2605.00950#pg2> The Hankel-DMD algorithm successfully separated the signal components, identifying the true structural natural frequency at 0.524 Hz, distinct from the 3P forcing <ref:2605.00950#pg2> This spectral selectivity ensures that the Digital Twin tracks the structural stiffness rather than simply monitoring the rotational speed of the turbine <ref:2605.00950#pg2> Furthermore, extending the rank to r = 90 was strictly necessary for real-time virtual sensing to resolve fine-scale spatial correlations <ref:2605.00950#pg2> Virtual Sensing and Prediction The framework employs a rolling horizon strategy that alternates between calibration and prediction for robust reconstruction over the full duration t0 → tn <ref:2605.00950#pg2> The reconstruction proceeds via a single vectorized matrix operation, replacing iterative time-stepping with efficient linear algebra <ref:2605.00950#pg2> The reconstructed trajectory Yˆ rec is computed through the formula Re(Φphys · diag(bk) · VW) <ref:2605.00950#pg2> This formulation replaces iterative time-stepping with efficient linear algebra, acting as a filter that continuously prevents drift <ref:2605.00950#pg2> Lyapunov Limit and Stability A critical finding is the identification of the prediction horizon, or the Lyapunov Limit, of the Digital Twin <ref:2605.00950#pg2> When the data assimilation interval was extended to Tup = 2.0 s, a marked degradation in performance was observed <ref:2605.00950#pg2> The logarithmic divergence of trajectories follows a linear trend with a slope of λmax ≈ 1.23, corresponding to a Lyapunov time (prediction horizon) of Tλ ≈ 0.81 s <ref:2605.00950#pg2> This theoretical limit aligns with the virtual sensing results, explaining the rapid degradation in accuracy when the open-loop update horizon (2.0 s) exceeds the valid chaotic prediction window <ref:2605.00950#pg2> The requirement for data assimilation updates at a frequency of approximately 1 Hz is highly practical for deployment <ref:2605.00950#pg2> This study demonstrates that the Hankel-DMD framework provides a robust, computationally efficient foundation for Floating Offshore Wind Turbine Digital Twins <ref:2605.00950#pg2> The model retains the spectral interpretability of physics-based models while bypassing the limitations of both opaque neural networks and rigid physical assumptions <ref:2605.00950#pg2> The framework successfully isolates true structural resonances from high-energy harmonic interference, such as rotor-induced 3P frequencies <ref:2605.00950#pg2> The results confirm that the Hankel formulation successfully isolates true structural resonances from high-energy harmonic interference, such as rotor-induced 3P frequencies <ref:2605.00950#pg2> The framework achieves Normalized Root Mean Square Errors (NRMSE) consistently below 1% for missing sensor reconstruction, even under active but unmeasured stochastic loading <ref:2605.00950#pg2> The inherent versatility of this structurally independent framework suggests a scalable pathway for Fleet-Level Monitoring <ref:2605.00950#pg2> The reproducible Python implementation of the Hankel-DMD / Koopman-Hankel digital twin framework is available at https://github.com/Mahdi22223/FOWT-Digital-Twin <ref:2605.00950#pg2> The archived software release is available via Zenodo: https://doi.org/10.5281/zenodo/20110415 <ref:2605.

Paper summary: Rosa: We're moving into what this paper is actually claiming about the Equation-Free Digital Twins for Nonlinear Structural Dynamics. Essentially, the thesis is that monitoring these high dimensional structures in extreme environments is limited by nonlinearities and complex forces, and traditional methods just don't cut it for real time.

Dev: The authors claim their rank optimized digital twin framework based on Koopman Operator Theory using Hankel-matrix embeddings and dynamic mode decomposition bridges the gap between high fidelity physics and real time processing.

Rosa: They achieve this by lifting the operational data into a linear invariant subspace, which lets the model reconstruct the structural state autonomously without needing mass or stiffness matrices beforehand.

Dev: The core objective here is to find a finite dimensional approximation of that Koopman operator, which governs how those sensor measurements evolve over time in R p.

Rosa: They are specifically looking for this operator A, and they use Hankel-matrix embeddings and dynamic mode decomposition to get that rank optimized structure.

Dev: This approach is supposed to eliminate the need for those upfront mass stiffness matrices while still keeping some physical interpretability in the results.

Taro: When I think about failure modes, this means we can get a representation of the structure that doesn't rely on perfect initial physics models to run in real time.

Rosa: Exactly. They are trying to find this operator A, and they use Hankel-matrix embeddings and dynamic mode decomposition to get that rank optimized structure.

Conclusion: Rosa: So, wrapping up the Equation-Free Digital Twins for Nonlinear Structural Dynamics paper by Mohammad Mahdi Abaei and the team, it shows that the Hankel-DMD framework is a robust and computationally efficient foundation for floating offshore wind turbine digital twins.

Dev: It successfully retains the spectral interpretability of physics based models while bypassing those limitations with both opaque neural networks and rigid physical assumptions.

Rosa: The key achievement here is that the model isolates true structural resonances from high energy harmonic interference, like rotor induced three power frequencies, and they got normalized root mean square errors below one point for missing sensor reconstruction even under active but unmeasured stochastic loading.

Dev: This structural independence suggests a scalable pathway for fleet level monitoring because it doesn't need those rigid physical assumptions to handle the nonlinearity.

Taro: The reproducible Python implementation of this Hankel-DMD Koopman Hankel digital twin framework is available at https://github.com/Mahdi22223/FOWT-Digital-Twin, and the archived software release is available via Zenodo at https://doi.org/ten point five two eight one/zenodo/twenty million one hundred ten thousand four hundred fifteen.

Rosa: It’s a really versatile approach, and it suggests a way forward for monitoring these massive structures in complex environments without getting stuck in the limitations of current techniques like standard subspace identification or traditional state space models.

Department of Mechanical Engineering, Marine Technology and Arctic Group, Aalto University · Independent Expert, Digital Twin Technology Ribe Denmark Institute for Risk and Reliability Leibniz University Hannover Germany

eess.SP, cs.LG, cs.SY, eess.SY

Submitted: 2026-05-01

Updated: 2026-11-11

Comments: Added more details of the theory as well as real experimental validation also included in the results. Now the manuscript is more defensible

Code: https://github.com/Mahdi22223/FOWT-Digital-Twin

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 87/100

The gist: Monitoring high-dimensional engineering structures in extreme environments is fundamentally limited not only by nonlinearity, but also by the convergence of non-stationary excitation…

Key concepts

Koopman Operator Theory
This theory transforms complex nonlinear systems into linear ones by mapping their evolution into a higher-dimensional functional space. The goal is to find a finite-dimensional approximation of this operator, which allows the system's behavior to be described using standard linear algebra techniques, making it easier to model and predict.
Hankel-Matrix Embeddings
These are mathematical tools used to embed time-series data into a higher dimension. By constructing Hankel matrices from sensor measurements, the framework creates a representation that captures the underlying dynamics of the system. This embedding is crucial for identifying patterns in high-dimensional, nonstationary data.
Dynamic Mode Decomposition (DMD)
DMD is an algorithm used to decompose complex time-series data into a set of dominant modes and their corresponding frequencies. In this context, it helps separate the structural vibrations from external forcing like rotor harmonics, allowing the system to identify the true natural frequencies of the structure.
Lyapunov Limit (Prediction Horizon)
This is a critical theoretical limit defining how long a Digital Twin can accurately predict future states based on current data. The study found this limit is short (around 0.81 seconds), indicating that extending the data update interval beyond this point causes the model's predictions to diverge rapidly, guiding practical deployment frequency.

Terminology

Summary

Monitoring high-dimensional engineering structures in extreme environments is fundamentally limited not only by nonlinearity, but also by the convergence of non-stationary excitation, nonlinear structural kinematics, and complex stochastic forcing <ref:2605. p. 000950#pg2> This paper introduces a rank-optimized Digital Twin framework based on Koopman Operator Theory, utilizing Hankel-matrix embeddings and dynamic mode decomposition to bridge the gap between high-fidelity physics and real-time edge processing <ref:2605. p. 000950#pg2> By lifting operational data into a linear invariant subspace, the model achieves an autonomous, input-blind reconstruction of the structural state without requiring a priori mass or stiffness matrices <ref:2605. p. 000950#pg2>

Framework Overview

The proposed framework is an equation-free digital twin based on Koopman Operator Theory that bridges high-fidelity physics and real-time edge processing <ref:2605.00950#pg2> It achieves this by lifting nonlinear dynamics into a higher-dimensional functional space where their evolution is linear <ref:2605. p. 200950#pg2> The core objective is to find a finite-dimensional approximation of the Koopman operator, denoted as A, which governs the time evolution of available sensor measurements yk in R p <ref:2605. p. 500950#pg2> This approach eliminates the need for a priori mass-stiffness matrices while maintaining physical interpretability <ref:2605. p. 300950#pg2>

How the training signal is built

The training dataset was generated using OpenFAST, modeling the NREL 5MW wind turbine on the OC3 Hywind Spar platform <ref:2605.00950#pg2> The simulation setup was explicitly designed to capture the chaotic, coupled nature of the offshore environment essential for identification <ref:2605. p. 300950#pg2> To ensure a robust training set, a randomized training ensemble was constructed by sampling from the global dataset, preventing overfitting to specific time-histories <ref:2605. p. 300950#pg2> The state vector is constructed by fusing two distinct physical quantities: absolute acceleration and bending moment (derived from strain) <ref:2605. p. 400950#pg2>

Where the gains come from

The rank-optimized manifold effectively unmixes structural resonances from deterministic 3P rotor harmonics, a task where standard subspace identification typically fails under colored noise <ref:2605.00950#pg2> The Hankel-DMD algorithm successfully separated the signal components, identifying the true structural natural frequency at 0.524 Hz, distinct from the 3P forcing <ref:2605. p. 000950#pg2> This spectral selectivity ensures that the Digital Twin tracks the structural stiffness rather than simply monitoring the rotational speed of the turbine <ref:2605. p. 000950#pg2> Furthermore, extending the rank to r = 90 was strictly necessary for real-time virtual sensing to resolve fine-scale spatial correlations <ref:2605. p. 300950#pg2>

Virtual Sensing and Prediction

The framework employs a rolling horizon strategy that alternates between calibration and prediction for robust reconstruction over the full duration t0 → tn <ref:2605.00950#pg2> The reconstruction proceeds via a single vectorized matrix operation, replacing iterative time-stepping with efficient linear algebra <ref:2605. p. 900950#pg2> The reconstructed trajectory Yˆ rec is computed through the formula Re(Φphys · diag(bk) · VW) <ref:2605. p. 100950#pg2> This formulation replaces iterative time-stepping with efficient linear algebra, acting as a filter that continuously prevents drift <ref:2605.00950#pg2>

Lyapunov Limit and Stability

A critical finding is the identification of the prediction horizon, or the Lyapunov Limit, of the Digital Twin <ref:2605.00950#pg2> When the data assimilation interval was extended to Tup = 2. p. 00 s, a marked degradation in performance was observed <ref:2605.00950#pg2> The logarithmic divergence of trajectories follows a linear trend with a slope of λmax ≈ 1. p. 323, corresponding to a Lyapunov time (prediction horizon) of Tλ ≈ 0. p. 081 s <ref:2605. p. 000950#pg2> This theoretical limit aligns with the virtual sensing results, explaining the rapid degradation in accuracy when the open-loop update horizon (2. p. 00 s) exceeds the valid chaotic prediction window <ref:2605. p. 000950#pg2> The requirement for data assimilation updates at a frequency of approximately 1 Hz is highly practical for deployment <ref:2605. p. 000950#pg2>

Conclusion

This study demonstrates that the Hankel-DMD framework provides a robust, computationally efficient foundation for Floating Offshore Wind Turbine Digital Twins <ref:2605.00950#pg2> The model retains the spectral interpretability of physics-based models while bypassing the limitations of both opaque neural networks and rigid physical assumptions <ref:2605. p. 300950#pg2> The framework successfully isolates true structural resonances from high-energy harmonic interference, such as rotor-induced 3P frequencies <ref:2605.00950#pg2> The results confirm that the Hankel formulation successfully isolates true structural resonances from high-energy harmonic interference, such as rotor-induced 3P frequencies <ref:2605. p. 000950#pg2> The framework achieves Normalized Root Mean Square Errors (NRMSE) consistently below 1% for missing sensor reconstruction, even under active but unmeasured stochastic loading <ref:2605.00950#pg2> The inherent versatility of this structurally independent framework suggests a scalable pathway for Fleet-Level Monitoring <ref:2605. p. 300950#pg2> The reproducible Python implementation of the Hankel-DMD / Koopman-Hankel digital twin framework is available at https://github. p. 5com/Mahdi22223/FOWT-Digital-Twin <ref:2605. p. 000950#pg2> The archived software release is available via Zenodo: https://doi. p. 9org/10.5281/zenodo/20110415 <ref:2605. p. 0

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Higher fidelity structural state estimation in real-time is achieved by lifting nonlinear dynamics into a linear invariant subspace using Koopman Operator Theory, allowing for an autonomous, input-blind reconstruction of the structural state without requiring a priori mass or stiffness matrices.

  2. The improved system can perform virtual sensing on sparse sensor arrays by constructing a sparse mode basis and using a rolling horizon strategy to dynamically update modal amplitudes from available data via the equation bk = Phi†sparseyobs k.

  3. The AI system can achieve spectral selectivity by employing a deterministic rank optimization strategy, which effectively unmixes structural resonances from deterministic 3P rotor harmonics, isolating true structural natural frequencies from colored noise that typically confuses standard subspace identification methods.

  4. The Digital Twin can provide a rigorous physical constraint for data assimilation by quantifying the physical Lyapunov time (≈ 1.0 s), thereby defining the fundamental predictability horizon required to maintain accuracy under unseen environmental forcing.

Sources

Related papers