Efficacy of the Weak Formulation of Sparse Nonlinear Identification in Predicting Vortex-Induced Vibrations

This study demonstrates that the weak formulation of sparse nonlinear identification (WSINDy) offers a robust, data-driven framework for accurately discovering and predicting the governing equations of vortex-induced vibrations, outperforming traditional models and standard SINDy in handling aperiodic dynamics.

The Gist

Imagine a long, flexible pole sticking out of a river. As the water rushes past, it doesn't just flow smoothly; it creates swirling eddies (vortices) that peel off the pole like leaves from a tree. These swirling eddies push and pull on the pole, causing it to shake, wobble, and dance. This is called Vortex-Induced Vibration (VIV).

If this shaking gets too strong, the pole can snap or break (like a bridge cable or an oil rig). Engineers need to predict exactly how this pole will dance so they can build it strong enough to survive.

The Problem: The "Guess-and-Check" Dance

For years, engineers tried to write math equations to predict this dance. They used simplified models, kind of like trying to describe a complex jazz improvisation using only a simple nursery rhyme.

• The Old Way: They assumed the water behaved in a very specific, predictable way. But in the real world, water is messy, chaotic, and full of surprises. When the water gets really turbulent, these old math models break down and give wrong answers.
• The New Idea: Instead of guessing the rules, why not just watch the dance and let a computer figure out the rules for us? This is called Data-Driven Discovery.

The Tools: SINDy and WSINDy

The researchers used two smart computer algorithms to learn the rules of the dance from data (either from computer simulations or real-world measurements).

1. SINDy (The "Sharp-Eyed" Detective):
   Imagine you are trying to figure out the recipe for a cake by tasting it. SINDy looks at the ingredients (data) and tries to find the simplest list of rules that explains the taste.

   • The Flaw: To figure out how the cake is changing, SINDy tries to measure the "speed" of the change at every single tiny moment. If your data has a little bit of noise (like a crumb on the scale or a shaky hand), SINDy gets confused. It thinks the noise is a real ingredient and adds weird, nonsensical rules to the recipe. It's like trying to measure the speed of a hummingbird by taking a photo with a shaky camera; the blur makes it look like the bird is teleporting.

2. WSINDy (The "Smooth-Operator" Detective):
   This is the star of the show. Instead of looking at the data point-by-point (which is shaky), WSINDy looks at the average behavior over a small stretch of time.

   • The Analogy: Imagine you are trying to hear a conversation in a noisy room.
     • SINDy tries to listen to every single word instantly. The background noise makes it hard to understand.
     • WSINDy listens to the whole sentence, then averages it out. The background noise gets smoothed away, and the main message (the physics) becomes clear.
   • By using "integrals" (mathematical averaging), WSINDy acts like a noise-canceling headphone for the data. It ignores the tiny, messy jitters and focuses on the big, important movements.

The Experiment: The Three Dance Floors

The researchers tested these detectives on three different "dance floors" (scenarios):

1. The Practice Floor (Perfect Data): They used a computer simulation where they knew the exact rules.

   • Result: Both detectives did a great job. They found the correct rules. This proved the method works in theory.

2. The Chaotic Floor (Real-World Data): They used data from a high-fidelity computer simulation of water flowing around a pole. This data was messy, noisy, and had complex, irregular movements (like a jazz solo).

   • Result: SINDy got confused by the noise. It started inventing fake rules to explain the static, leading to a model that would eventually break.
   • WSINDy stayed calm. It filtered out the noise and found the true, simple rules governing the dance, even when the movement was irregular and unpredictable.

3. The Full Orchestra (Reconstructing the Flow): They tried to use these rules not just to predict the pole's movement, but to reconstruct the entire swirling water pattern behind it.

   • Result: This was very hard (like trying to describe a whole symphony with just a few notes). Both methods struggled a bit, but WSINDy still did a better job of capturing the "soul" of the flow without getting lost in the details.

The Big Takeaway

The paper concludes that WSINDy is a superior tool for understanding complex fluid problems.

• Why it matters: In the real world, data is never perfect. It's always noisy. If you use the old method (SINDy), you might build a bridge based on a model that looks good on paper but fails in a storm because it was confused by the noise.
• The Future: By using the "smooth-operator" approach (WSINDy), engineers can build better models for wind turbines, offshore oil rigs, and bridges. They can predict how these structures will behave in chaotic storms, leading to safer and more efficient designs.

In short: When trying to understand a messy, noisy system, don't just look at the individual, shaky steps. Step back, smooth out the motion, and you'll see the true rhythm of the dance.

Technical Summary

1. Problem Statement

Vortex-Induced Vibrations (VIV) are complex fluid-structure interactions (FSI) where bluff bodies (like cylinders) experience self-sustained oscillations due to vortex shedding. Predicting VIV is critical for structural integrity (preventing fatigue) and energy harvesting.

• Limitations of Traditional Models: Existing reduced-order models (e.g., wake-oscillator models based on Van der Pol equations) rely on empirical fitting and specific coupling assumptions. They often fail to capture complex 3D dynamics, multi-frequency spectral content, and regime transitions (lock-in to desynchronization) across varying Reynolds numbers and mass-damping parameters.
• Limitations of Standard Data-Driven Methods: While Sparse Identification of Nonlinear Dynamics (SINDy) offers a way to discover governing equations directly from data, the standard "strong-form" approach relies on numerical differentiation. This makes it highly sensitive to noise, measurement errors, and broadband fluctuations common in VIV data (especially from Computational Fluid Dynamics or experiments), leading to spurious terms and unstable models.

2. Methodology

The study conducts a systematic comparative assessment between Standard SINDy and its Weak Formulation (WSINDy) for discovering governing equations of VIV systems.

• Data Sources:
  1. Synthetic Data: Generated from a coupled wake-oscillator model (Van der Pol oscillator coupled with a structural mass-spring-damper) across reduced velocities ($U_r$) from 1 to 20. This provides "ground truth" equations.
  2. High-Fidelity CFD Data: Simulations of an elastically mounted 1-DoF circular cylinder at $Re=150$ across $U_r \in [3.0, 5.5]$, covering pre-lock-in, lock-in, and post-lock-in regimes.
• Identification Frameworks:
  • Standard SINDy: Uses Sequential Thresholded Least Squares (STLSQ) to solve $\dot{X} = \Phi(X)\Xi$. It requires computing time derivatives ($\dot{X}$) via finite differences, making it prone to noise amplification.
  • WSINDy: Reformulates the problem using an integral (weak) form: $\int \phi'(t)X(t) + \phi(t)\sum w_j f_j(X(t)) dt = 0$. It replaces derivatives with weighted integrals against smooth test functions ($\phi$), acting as a natural low-pass filter. It uses the Sparse Relaxed Regularized Regression (SR3) optimizer.
• State Variables:
  • System Level: Structural displacement ($y$) and wake parameter ($q$, derived from lift force).
  • Flow Field Level: Temporal coefficients of Proper Orthogonal Decomposition (POD) modes extracted from CFD velocity fields to reconstruct near-wake dynamics.
• Validation Metrics: Trajectory matching (time series), amplitude response curves, frequency spectra (FFT and Wavelet), and physical interpretability of the discovered terms.

3. Key Contributions

1. First Systematic Comparison: This is the first study to rigorously compare Standard SINDy and WSINDy specifically for VIV governing equation discovery across all canonical response regimes (aperiodic, lock-in, desynchronization).
2. Full-System Coupled Identification: Unlike previous studies that identified wake or structural equations separately or within pre-assumed structures, this work identifies the fully coupled structural-wake dynamics without a priori structural assumptions.
3. Demonstration of Weak Form Superiority: The study proves that WSINDy is significantly more robust than Standard SINDy when dealing with the broadband, noisy, and aperiodic dynamics characteristic of VIV, particularly in the pre-lock-in regime.
4. POD-SINDy Extension: The paper extends sparse identification to high-dimensional flow fields using POD, analyzing the trade-offs between library complexity, dimensionality, and numerical stability.

4. Key Results

A. Performance on Synthetic Data (Wake-Oscillator)

• Standard SINDy: Successfully recovered the governing equations with high fidelity in the lock-in regime. However, performance degraded in the post-lock-in regime ($U_r \ge 15$) due to increased dynamical complexity and amplitude modulation, leading to phase drift.
• Implication: Even with clean data, standard SINDy struggles when dynamics become highly non-periodic.

B. Performance on High-Fidelity CFD Data

• Aperiodic Regime ($U_r = 3.0, 3.5$):
  • Standard SINDy: Failed to track the true dynamics. The noise in the lift coefficient signal caused derivative estimation errors, leading to significant phase and amplitude deviations. The identified models diverged quickly.
  • WSINDy: Demonstrated markedly superior reconstruction fidelity. The integral formulation filtered out high-frequency noise, allowing the model to accurately track both structural displacement and wake variables.
• Lock-in Regime ($U_r = 4.0 - 5.5$):
  • Both methods performed well in tracking periodic, single-frequency oscillations.
  • WSINDy produced sparser and more physically coherent equations. Standard SINDy introduced spurious high-order nonlinear terms to compensate for noise, whereas WSINDy retained only the essential physical terms (e.g., Van der Pol damping).
• Stability Analysis: WSINDy remained stable across a broader range of sparsity thresholds ($\lambda$) compared to Standard SINDy, which frequently diverged at low or high thresholds due to overfitting or underfitting.

C. POD-Based Flow Reconstruction

• When applied to 10 POD modes (temporal coefficients) to reconstruct the flow field:
  • Linear Libraries: Both methods captured qualitative flow topology, but WSINDy provided better quantitative agreement.
  • Nonlinear Libraries: Attempting to use 2nd or 3rd-degree polynomial libraries for the 10-variable system resulted in instability for both methods. This highlights that the primary barrier is not library expressiveness but the sensitivity of high-dimensional systems to noise and the lack of physical constraints on inter-modal coupling.

5. Significance and Conclusion

• Robustness to Noise: The study conclusively demonstrates that the weak formulation (WSINDy) is essential for identifying governing equations from real-world or CFD-derived VIV data, where noise and broadband fluctuations render standard differentiation unreliable.
• Physical Interpretability: WSINDy yields models that are not only accurate in trajectory prediction but also physically interpretable, avoiding the "term proliferation" seen in Standard SINDy.
• Future Directions: The authors suggest that future work should focus on noise-aware sparse regression, physics-informed regularization to constrain the search space, and extending these methods to 3D geometries and tandem cylinder arrangements.

Summary: The paper establishes WSINDy as a superior, robust alternative to standard SINDy for VIV modeling, enabling the discovery of interpretable, predictive governing equations directly from noisy, high-fidelity data without requiring prior knowledge of the system's structure.