Data-Driven Modal Decomposition Analysis of Unsteady Flow in a Multi-Stage Turbine

This study applies Proper Orthogonal Decomposition (POD) and four Dynamic Mode Decomposition (DMD) variants to analyze unsteady flow in a 1.5-stage axial turbine, revealing that while specific DMD methods achieve reconstruction accuracy comparable to POD and better capture the system's true dynamic frequencies, both approaches identify dominant modes whose characteristics correlate with the turbine's adiabatic efficiency across different stator clocking configurations.

The Gist

Imagine a multi-stage turbine (like the heart of a jet engine) as a busy, chaotic dance floor. Inside, giant blades spin at thousands of revolutions per minute, creating a swirling, unpredictable storm of air pressure. Engineers need to understand this storm to build better, more efficient engines, but the data is so massive and complex that it's like trying to understand a hurricane by looking at every single raindrop individually.

This paper is about two different "smart cameras" (mathematical tools) that try to take that chaotic storm and break it down into simple, understandable patterns. The authors are comparing these two cameras to see which one tells the truest story about how the air moves.

Here is the breakdown of their findings using everyday analogies:

1. The Two Cameras: POD vs. DMD

The researchers used two main tools to analyze the airflow:

• POD (Proper Orthogonal Decomposition): Think of this as a Photo Album. It takes all the snapshots of the storm and finds the "average" picture and the most common "variations" from that average. It's great at compressing the data (making the file size smaller) so you can see the big picture. However, it treats time like a static list of photos; it doesn't really tell you how the storm evolves from one second to the next. It's like looking at a collage of a dancer's moves without seeing the flow of the dance.
• DMD (Dynamic Mode Decomposition): Think of this as a Movie Reel. It doesn't just look at the shapes; it looks at how those shapes move and change over time. It identifies specific "characters" (modes) in the storm, tells you how fast they spin, and whether they are fading away or getting stronger. It captures the rhythm of the dance.

2. The Experiment: Sorting the Noise

The researchers had a huge pile of data (snapshots of air pressure) from a 1.5-stage turbine. They tried to rebuild the original storm using only a few of these "characters" (modes).

• The Problem with Frequency Sorting: They tried sorting the DMD characters by how fast they vibrated (frequency). This was like trying to organize a library by the color of the book covers. It didn't work well because the most important "characters" in the storm weren't necessarily the ones vibrating the fastest. The reconstruction was messy and inaccurate.
• The Winners: They found that sorting by Amplitude (how loud/big the character is) or using a special "energy score" (Tissot criterion) worked perfectly. These methods, along with a "sparsity" method (which picks the fewest, most important characters), rebuilt the storm almost as perfectly as the Photo Album (POD) did.

3. What Did They Find?

• The "Mean" and the "Rhythm": Both cameras agreed on the first few patterns.
  • The 1st Mode was just the "average" air pressure (the calm background).
  • The 2nd and 3rd Modes were the real stars. They captured the main "beat" of the turbine, caused by the rotor blades passing by. These two modes are like a pair of dancers moving in perfect sync (one is the real part, one is the imaginary part of the same rhythm).
• The Truth About Time: Here is the big revelation. The POD camera was great at recreating the look of the storm, but it lied about the time. It couldn't tell you the true frequency of the beat. It was like a photo album that showed the dancer in all the right poses but made it look like they were moving at random speeds.
  • The DMD camera, however, told the truth. It showed that the storm is mostly driven by a steady, neutral rhythm (the rotor passing frequency) with a few fading echoes. It correctly identified the "heartbeat" of the machine.

4. The Clocking Effect: Tuning the Engine

The researchers also played with "clocking." Imagine the turbine has two rows of stationary blades (stators) and one row of spinning blades (rotor). "Clocking" is simply rotating the top row of stationary blades slightly left or right, like changing the time on a clock face.

• The Discovery: They found that when the engine runs most efficiently (highest "adiabatic efficiency"), the "dance" of the air becomes more energetic.
• The Connection: The configurations that made the engine run best were the ones where the 2nd and 3rd DMD modes (the main dancers) were the loudest and most active.
• The Analogy: It's like a choir. When the choir sings the most beautifully (highest efficiency), the lead singers (the dominant modes) are singing the loudest and most clearly. If you can measure how loud those lead singers are, you know how well the engine is performing.

The Bottom Line

This paper teaches us that while old-school methods (POD) are good for compressing data, Dynamic Mode Decomposition (DMD) is the superior tool for understanding how a machine works over time.

By using DMD, engineers can now look at the "rhythm" of the airflow and predict how changing the blade positions (clocking) will affect the engine's efficiency. It turns a chaotic, unseeable storm of air into a clear, rhythmic dance that can be optimized for better performance.

Technical Summary

1. Problem Statement

The paper addresses the challenge of analyzing and understanding unsteady flow phenomena in multi-stage turbomachines, specifically a 1.5-stage subsonic axial turbine. While the Navier-Stokes equations provide an accurate physical model, their nonlinearity makes obtaining fast, accurate numerical solutions difficult. Traditional methods often struggle to efficiently extract dominant flow structures and dynamic features from high-dimensional flow data.

The study focuses on:

• Unsteady Flow Dynamics: Analyzing the complex interactions (wake effects and potential disturbances) between blade rows, particularly in the downstream stator (Stator 2), which significantly impacts overall turbine performance.
• Methodological Comparison: Evaluating the effectiveness of two primary data-driven modal decomposition techniques—Proper Orthogonal Decomposition (POD) and Dynamic Mode Decomposition (DMD)—in reconstructing flow fields and identifying dynamic features.
• Mode Selection: Determining the most suitable criteria for selecting dominant modes in DMD, as DMD lacks a universal sorting criterion unlike POD.
• Performance Correlation: Investigating the relationship between specific flow modes and the turbine's adiabatic efficiency under different stator clocking configurations.

2. Methodology

The authors employed a combination of computational fluid dynamics (CFD) and advanced data analysis algorithms:

• Flow Generation:
  • Geometry: A 1.5-stage axial turbine (Stator 1, Rotor, Stator 2) with specific blade profiles (Traupel for stators, modified VKI for rotor).
  • Simulation: Unsteady Reynolds-Averaged Navier-Stokes (URANS) equations were solved using an in-house CFD code with a second-order cell-centered finite-volume method.
  • Data: 120 time snapshots of the flow field (covering 4 rotor passing periods) were extracted from the downstream stator (Stator 2) at 50% span.
• Modal Decomposition Algorithms:
  • POD: Applied via Singular Value Decomposition (SVD) to extract orthogonal spatial modes based on energy content (singular values).
  • DMD: Applied to extract dynamic modes based on linear evolution between snapshots. The study tested four DMD variants:
    1. Amplitude Criterion: Sorting modes by amplitude modulus ($|\alpha_k|$).
    2. Frequency Criterion: Sorting modes by angular frequency ($|\omega_k|$).
    3. Tissot Criterion: Sorting by amplitude weighted by growth rate ($E_k$).
    4. Sparsity-Promoting DMD (SP-DMD): Using a penalty function to identify optimal non-zero mode combinations.
• Implementation: A parallel code using MPI and TSQR decomposition was developed to handle the high dimensionality of the flow data ($m \approx 1.4$ million grid points).

3. Key Contributions

• Comparative Evaluation: A rigorous comparison of POD and four DMD variants specifically for multi-stage turbine flows, a domain where such comparisons are rare.
• Validation of Mode Selection Criteria: Identification of the Amplitude, Tissot, and SP-DMD criteria as effective for this problem, while demonstrating the Frequency Criterion is unsuitable for capturing dominant unsteady features in this context.
• Dynamic vs. Statistical Insight: A critical distinction is drawn between POD (which provides optimal reconstruction but misrepresents true dynamic frequencies) and DMD (which captures true temporal evolution and single-frequency dynamics).
• Performance-Mode Correlation: Establishing a direct link between the strength of specific low-order modes and the turbine's adiabatic efficiency across different clocking configurations.

4. Key Results

A. Reconstruction Accuracy

• POD vs. DMD: POD provides the most accurate reconstruction by definition (minimizing Frobenius norm error). However, DMD methods based on the Amplitude, Tissot, and SP-DMD criteria achieve reconstruction accuracy comparable to POD (residuals $< 10^{-4}$ with 7 modes).
• Frequency Criterion Failure: The DMD method using the frequency criterion resulted in significantly higher errors (one order of magnitude larger). This is because it selected modes with high decay rates that do not persist in the globally periodic flow, failing to capture the dominant physics.
• Prediction Capability: Unlike POD, DMD successfully predicted future flow states. With 7 modes, both POD and valid DMD variants accurately reconstructed and predicted pressure fluctuations, including high-frequency components and decay behaviors.

B. Mode Shapes and Dynamics

• Dominant Modes: The first mode for both methods represents the time-averaged flow. The 2nd and 3rd modes (a complex-conjugate pair in DMD) capture the dominant pressure fluctuation structures driven by the rotor passing frequency (RPF).
• Spatial Similarity: POD and DMD spatial modes showed strong similarity. The 2nd POD mode closely matched the 3rd DMD mode, and the 3rd POD matched the 2nd DMD (with sign reversal).
• Dynamic Features:
  • DMD: Correctly identified that the flow is governed by neutral modes (zero growth rate) at the RPF and its harmonics (2x RPF). It accurately captured the phase shift between real and imaginary parts of conjugate pairs.
  • POD Limitation: While POD reconstructed the flow well, its time coefficients were purely periodic and misrepresented the fundamental frequencies. POD modes contained mixed frequencies (superimposed harmonics), failing to isolate the true dynamic components of the system.

C. Stator Clocking and Efficiency

• Clocking Effect: Different stator clocking configurations ($CC_0$ to $CC_4$) altered the aerodynamic performance.
• Correlation: A strong correlation was found between adiabatic efficiency and mode strength.
  • Configurations with higher efficiency corresponded to larger amplitudes (modulus) of the 2nd and 3rd DMD mode pairs and the 2nd POD mode.
  • This suggests that higher efficiency is associated with larger, more coherent pressure fluctuation amplitudes in the dominant modes, rather than their suppression.
• Higher Modes: No significant correlation was found between higher-order modes (4th and beyond) and aerodynamic efficiency.

5. Significance

This study provides critical insights for the design and optimization of multi-stage turbines:

1. Method Selection: It validates that DMD (specifically using Amplitude, Tissot, or SP-DMD criteria) is a superior tool for understanding the physics of unsteady flows in turbines compared to POD, as it correctly identifies dynamic frequencies and growth rates.
2. Design Guideline: The finding that higher efficiency correlates with larger dominant mode amplitudes challenges the intuition that suppressing fluctuations always improves performance. Instead, it suggests that optimizing the coherence and amplitude of specific rotor-stator interaction modes is key to efficiency.
3. Reduced-Order Modeling (ROM): The results confirm that a small number of modes (approx. 7) are sufficient to reconstruct and predict unsteady flows in multi-stage turbines, enabling efficient ROMs for real-time control or rapid design iteration.
4. Limitation of POD: The paper highlights a fundamental limitation of POD in dynamic analysis: while excellent for data compression, it can obscure the true dynamic nature (frequencies and stability) of the system, making DMD essential for dynamic feature extraction.