TL;DR
This paper introduces a novel neural network model to predict the onset of laminar-to-turbulent transition in boundary layers, specifically addressing the challenge posed by 2D surface defects like gaps and steps. By training on a comprehensive database of simulations, their model accurately predicts the destabilizing effect of these defects, offering a more generalized and efficient approach than traditional methods.
What Problem Does This Paper Address?
Predicting when a smooth, laminar boundary layer transitions to a turbulent one is crucial for aerodynamic design, as turbulence significantly increases drag. While methods like the "eN method" are effective for smooth surfaces, they struggle to account for the complex influence of surface imperfections such as rivets, gaps, or steps, which are common in real-world aircraft manufacturing. These defects can prematurely trigger turbulence, negating the benefits of laminar flow strategies. Existing empirical correlations for defects are often limited to specific geometries and conditions, necessitating a more adaptable and predictive tool.
Methods
The core of this research lies in a neural network (NN) approach. The authors first generate a comprehensive database of boundary layer stability computations using the ONERA code PIMS2D. These computations involve solving the linearized Navier-Stokes (LNS) equations for various defect geometries (defined by heights $h_1$, $h_2$, and width $b^*$) and flow conditions (characterized by the Reynolds number $Re_{\delta_1,d}$). The output of these simulations is the $\Delta N$ value, which quantifies the shift in the transition point caused by the defect.
This generated data then serves as training material for a neural network. The NN takes the defect's geometric parameters ($h_1^*$, $h_2^*$, $b^*$) and the aerodynamic parameter ($Re_{\delta_1,d}$) as input and is trained to predict two key outputs: $\Delta N_{max}$ (the maximum $\Delta N$ value) and $\Delta N_{far}$ (the $\Delta N$ value far downstream of the defect). The authors explore different NN architectures, employing a rectified linear unit (ReLU) activation function and minimizing the Mean Squared Error (MSE) loss function. The model's performance is evaluated using the Mean Absolute Error (MAE).
Datasets
The primary dataset used in this study was generated by the authors using the PIMS2D computational fluid dynamics code. This dataset comprises 750 observations, covering a range of geometric parameters for surface defects (e.g., $Re_{\delta_1,d}$ from 901.53 to 1999.41, $h_1^*$ from 0.10 to 3, $h_2^*$ from 0 to 2.90, $b^*$ from 0.52 to 14.98) and their corresponding $\Delta N_{max}$ and $\Delta N_{far}$ values. This dataset was split into training (80%) and validation (20%) sets.
The authors also validate their model against existing experimental data from Crouch et al. [4] and Methel et al. [10], as well as empirical correlations from Wang and Gaster [16] and Hildebrand et al. [9].
Key Insights and Ideas
- Generalization Beyond Empirical Limits: The NN approach moves beyond the limitations of specific empirical correlations for defects, offering a more generalized prediction capability for a wider range of gap-like surface irregularities.
- Unified Parameterization: By using a consistent set of geometric parameters ($h_1^*$, $h_2^*$, $b^*$) and an aerodynamic parameter ($Re_{\delta_1,d}$), the study standardizes the characterization of various defect types, including backward-facing steps (BFS) and cavities.
- Strong Correlation Between $\Delta N_{max}$ and $\Delta N_{far}$: The analysis reveals a strong linear relationship between $\Delta N_{max}$ and $\Delta N_{far}$, suggesting that predicting one might allow for accurate estimation of the other, potentially simplifying future models.
- Defect Geometry Dominance: The study highlights how defect geometry, particularly the difference in heights ($\Delta h = |h_1^* - h_2^*|$), plays a significant role in destabilizing the boundary layer, with BFS-like configurations (large $\Delta h$) being more impactful.
Key Results
- The developed neural network models achieve a Mean Absolute Error (MAE) of approximately $8 \times 10^{-2}$ on the $\Delta N$ prediction, indicating a robust prediction of surface irregularity criticality.
- The NN model shows good agreement with experimental data from Crouch et al. [4] for gap configurations, even for parameters outside the direct training range, demonstrating its generalization capabilities.
- Comparisons with Methel et al. [10] for gap configurations with wall suction also show reasonable agreement between experimental $\Delta N_{Exp}$ and the NN's predicted $\Delta N_{far}$.
- Validation against Wang and Gaster [16] for BFS configurations shows promising results, with the NN predictions aligning well with experimental data and empirical correlations, especially within the trained parameter space.
Limitations and Future Work
- Limited Range of Training Data: While the NN shows good generalization, its accuracy is best within the range of parameters it was trained on. Performance may degrade for defects significantly outside this range (e.g., very large widths or heights).
- Focus on 2D Defects: The current model is specifically designed for 2D surface defects. Extending it to 3D imperfections would be a significant next step.
- Compressibility and Pressure Gradients: The current study focuses on incompressible flow and assumes a Blasius boundary layer. Future work could incorporate the effects of compressibility and non-zero pressure gradients.
- Experimental Validation Expansion: While validated against existing data, further experimental validation across a broader spectrum of defect types and flow conditions would strengthen the model's credibility.
- Further Exploration of $\Delta N_{max}$ vs. $\Delta N_{far}$: The strong correlation between $\Delta N_{max}$ and $\Delta N_{far}$ warrants further investigation to potentially simplify the prediction task by focusing on just one output.
Summary updated