Rapid Prediction of Ice Accretion on Swept Wings Based on Proper Orthogonal Decomposition and Surrogate Modelling

Juan Du, Qiao‐Nan Guo, Yinghong Yue, Yao Ma, Hao ChengJournal of Applied Fluid Mechanics 2025 · DOI

This academic paper, published in the Journal of Applied Fluid Mechanics, presents a significant step forward in aircraft safety and aerodynamic design. Researchers from the Civil Aviation Flight University of China have developed a way to predict how ice builds up on complex swept wings in seconds, rather than hours.

TL;DR

By combining Proper Orthogonal Decomposition (POD) with Kriging surrogate modeling, researchers created a method to predict 3D ice accretion on swept wings three orders of magnitude faster than traditional simulations. The model achieves over 94% similarity for rime ice and 89% for glaze ice compared to high-fidelity numerical results.

What Problem Does This Paper Address?

Aircraft icing is a major flight safety hazard, causing increased drag and decreased lift. While straight-wing icing is well-studied, swept wings (common on modern jets) introduce complex 3D airflow, including spanwise pressure gradients and wingtip vortices, that make ice prediction much harder.

The gold standard for prediction is Computational Fluid Dynamics (CFD) combined with thermodynamic icing models. However, these simulations are "computationally intensive"—a single condition can take hours or even days to solve. This makes them impractical for real-time cockpit alerts or large-scale design optimizations where thousands of flight conditions must be tested.

Methods

The authors propose a "surrogate modelling" approach. Instead of solving the physics from scratch every time, they train a mathematical model to "guess" the result based on previous data. Their workflow involves:

  1. Data Generation: They used Optimized Latin Hypercube Sampling (OLHS) to select 120 representative icing conditions (varying speed, temperature, altitude, etc.) and ran high-fidelity simulations to create a training dataset.
  2. Coordinate Transformation: To simplify the math, they mapped 2D Cartesian coordinates $(x, y)$ to a 1D "ice height" $(\eta)$ along the wing's arc length $(\xi)$. This dimensionality reduction is crucial for model efficiency.
  3. POD (Proper Orthogonal Decomposition): This technique extracts the most essential "features" or basis functions from the icing data, effectively compressing complex shapes into a few key variables.
  4. Kriging Interpolation: A surrogate model was trained to map meteorological/flight parameters to the POD coefficients.
  5. 3D Reconstruction: The model predicts ice shapes at several 2D cross-sections along the wing and then uses uniform interpolation to "stitch" them into a full 3D geometry.

Datasets

  • Training/Testing Data: 120 icing conditions generated via numerical simulation based on FAA/EASA airworthiness regulations (FAR Part 25 Appendix C).
  • Validation Benchmarks: The model was validated against experimental ice shape data from the NASA Lewis Research Center (Shin and Bond, 1992) using the standard NACA 0012 airfoil.

Key Insights and Ideas

  • The Power of Dimensionality Reduction: By converting 2D shapes into 1D ice height data, the surrogate model avoids the "curse of dimensionality," allowing it to capture relationships between weather and ice shape much more accurately with fewer samples.
  • Linear vs. Nonlinear Behavior: The model performs exceptionally well on rime ice (which freezes instantly and has a smoother profile) because its growth is relatively linear. Glaze ice (which involves liquid runoff and "horn" formations) is more nonlinear and harder to predict, yet the model still maintains high accuracy.
  • Sensitivity Analysis: The study confirmed that ambient temperature is the primary driver of the type of ice (rime vs. glaze), while Liquid Water Content (LWC) and flight speed primarily dictate the rate of accumulation and the thickness of the ice "horns."

Key Results

  • Massive Speedup: Prediction time for a single condition was reduced from over 10,000 seconds (nearly 3 hours) to less than 10 seconds.
  • High Fidelity: The similarity score ($Par$) between the surrogate prediction and the full simulation exceeded 94% for rime ice and 89% for glaze ice.
  • 3D Accuracy: When reconstructed into a full 3D swept wing, the overall similarity remained above 90%, proving that interpolating between 2D slices is a valid shortcut for complex wing geometries.

Limitations and Future Work

While the model is incredibly fast, it tends to "simplify" or smooth out the most intricate geometric details of glaze ice horns. The authors note that the simulated ice shapes sometimes show more complex curvature than the surrogate model can currently capture.

Future Work suggested by the paper includes:

  • Increasing the density of the grid nodes on the airfoil surface to better capture "ice-corner" features.
  • Expanding the training dataset specifically for glaze ice conditions to better handle its nonlinear characteristics.
  • Integrating this rapid prediction tool into real-time aircraft anti-icing control systems.

Summary updated

Abstract

Numerical simulations of three-dimensional airfoil icing are computationally intensive, with icing complexities on swept wings surpassing those on straight wings. To enable rapid and accurate ice formation predictions on swept wings, this study proposes a prediction methodology integrating proper orthogonal decomposition (POD) and Kriging surrogate modelling. This approach incorporates key physical parameters influencing ice formation, including flight altitude, flight speed, ambient temperature, liquid water content, and median volume diameter. First, an optimized Latin hypercube sampling method (OLHS) was employed to generate 120 icing conditions under both continuous and intermittent maximum icing scenarios. Numerical simulations were then conducted to establish an icing dataset, which was subsequently transformed into one-dimensional ice height data for various two-dimensional airfoil sections. Next, surrogate models for two-dimensional airfoils were developed using POD and Kriging interpolation to establish relationships between meteorological and flight conditions and the corresponding icing shapes. Finally, three-dimensional ice geometries were reconstructed through uniform interpolation of multiple two-dimensional icing profiles. Validation results demonstrated a strong agreement between surrogate model predictions and numerical simulations, enabling rapid and accurate real-time ice shape estimations across various conditions. The predicted ice shape similarity exceeded 94% for rime ice and 89% for glaze ice. This methodology provides valuable insights for aircraft anti-icing and de-icing design while also contributing to the development of optimized ice-tolerant aerodynamic strategies.
BibTeX
@article{du2025rapid,
  author = {Juan Du and Qiao‐Nan Guo and Yinghong Yue and Yao Ma and Hao Cheng},
  title = {{Rapid Prediction of Ice Accretion on Swept Wings Based on Proper Orthogonal Decomposition and Surrogate Modelling}},
  journal = {Journal of Applied Fluid Mechanics},
  year = {2025},
  doi = {10.47176/jafm.18.8.3278},
}
Figure