From industry-wide parameters to aircraft-centric on-flight inference: improving aeronautics performance prediction with machine learning

Florent Dewez, Benjamin Guedj, Vincent Vandewalle

Abstract

Aircraft performance models play a key role in airline operations, especially in planning a fuel-efficient flight. In practice, manufacturers provide guidelines which are slightly modified throughout the aircraft life cycle via the tuning of a single factor, enabling better fuel predictions. However this has limitations, in particular they do not reflect the evolution of each feature impacting the aircraft performance. Our goal here is to overcome this limitation. The key contribution of the present article is to foster the use of machine learning to leverage the massive amounts of data continuously recorded during flights performed by an aircraft and provide models reflecting its actual and individual performance. We illustrate our approach by focusing on the estimation of the drag and lift coefficients from recorded flight data. As these coefficients are not directly recorded, we resort to aerodynamics approximations. As a safety check, we provide bounds to assess the accuracy of both the aerodynamics approximation and the statistical performance of our approach. We provide numerical results on a collection of machine learning algorithms. We report excellent accuracy on real-life data and exhibit empirical evidence to support our modelling, in coherence with aerodynamics principles.

Introduction

In the aviation industry, fuel efficiency is paramount. However, an aircraft’s performance is not static; it degrades over its lifecycle due to engine wear, surface impurities, and structural deformations. Currently, airlines rely on manufacturer-provided guidelines adjusted by a single, coarse "performance index." This approach lacks granularity, failing to dissociate different physical factors affecting performance. Furthermore, critical metrics like the coefficients of drag ($C_D$) and lift ($C_L$) cannot be measured directly by onboard sensors. This paper introduces a framework to move from generic models to aircraft-specific models powered by machine learning, leveraging data from Quick Access Recorders (QAR) to infer these latent variables.

Methodology

The authors propose a hybrid approach that blends physics-based modeling with machine learning to create a "digital twin" of an individual aircraft’s performance.

Physical Approximation

Since $C_D$ and $C_L$ are not directly recorded, the authors use flight dynamics equations (Newton’s second law) and established fuel consumption models to calculate "approximated" values. This step acts as a bridge, using physics to create surrogate labels for training.

Dependency graph illustrating how recorded flight parameters (e.g., velocity, mass, fuel flow) are used to approximate latent aerodynamic coefficients.

Machine Learning Framework

Using these approximations as targets, several regression models—including Polynomial Regression, Random Forests, and Gradient Tree Boosting (LightGBM)—are trained to predict $C_D$ and $C_L$ based on the angle of attack ($\alpha$) and Mach number ($M$). The study utilizes a private dataset of 423 flights from a single airliner, with data denoised via smoothing splines and filtered for steady-state cruise conditions.

Error Bounding

A key theoretical contribution is the derivation of mathematical bounds for the "total error," which sums the physical approximation error and the statistical learning error. This provides a necessary safety check for industrial application.

Table mapping abstract statistical variables to the specific aeronautic parameters used in the error bounding framework.

Key Results

  • High Accuracy in Cruise: For the drag coefficient ($C_D$), Gradient Tree Boosting and Random Forests achieved a Mean Absolute Percentage Error (MAPE) of approximately 4.29%.
  • Precision in Lift: For the lift coefficient ($C_L$), simpler models like Polynomial Regression performed exceptionally well, with a MAPE as low as 1.69%.
  • Physical Consistency: The models successfully captured expected aerodynamic trends, showing that drag and lift increase with the angle of attack and Mach number, validating that the ML models learned real physics.
  • Robustness: The calculated bounds for total error remained low (e.g., ~9.67% relative error for $C_D$), suggesting the approach is robust enough for practical flight planning.

Conclusion

This research demonstrates that machine learning can effectively individualize aircraft performance models by leveraging existing flight data. While the models perform excellently during cruise, future work will focus on improving stability during the climb phase, where fluctuating variables like Mach number introduce more noise. This methodology offers a path toward real-time, on-flight fuel optimization and more precise aeronautic engineering.

Summary updated

Abstract

Aircraft performance models play a key role in airline operations, especially in planning a fuel-efficient flight. In practice, manufacturers provide guidelines which are slightly modified throughout the aircraft life cycle via the tuning of a single factor, enabling better fuel predictions. However this has limitations, in particular they do not reflect the evolution of each feature impacting the aircraft performance. Our goal here is to overcome this limitation. The key contribution of the present article is to foster the use of machine learning to leverage the massive amounts of data continuously recorded during flights performed by an aircraft and provide models reflecting its actual and individual performance. We illustrate our approach by focusing on the estimation of the drag and lift coefficients from recorded flight data. As these coefficients are not directly recorded, we resort to aerodynamics approximations. As a safety check, we provide bounds to assess the accuracy of both the aerodynamics approximation and the statistical performance of our approach. We provide numerical results on a collection of machine learning algorithms. We report excellent accuracy on real-life data and exhibit empirical evidence to support our modelling, in coherence with aerodynamics principles.
BibTeX
@inproceedings{dewez2020from,
  author = {Florent Dewez and Benjamin Guedj and Vincent Vandewalle},
  title = {{From industry-wide parameters to aircraft-centric on-flight inference: improving aeronautics performance prediction with machine learning}},
  booktitle = {Data-Centric},
  year = {2020},
  doi = {10.1017/dce.2020.12},
}
Figure