Adding Flight Mechanics to Flight Loads Surrogate Model using Multi-Output Gaussian Processes

Ankit Chiplunkar, Emmanuel Rachelson, Michele Colombo, Joseph Morlier2016 · DOI

TL;DR

This paper introduces a novel method for building more accurate and physically consistent surrogate models of aircraft flight loads by incorporating known physics-based relationships between different load outputs. By leveraging multi-output Gaussian Processes, the authors demonstrate improved prediction capabilities, especially when dealing with limited or missing data, by effectively sharing information across related measurements.

What Problem Does This Paper Address?

Accurately predicting aircraft flight loads is crucial for design, certification, and maintenance, but experimental measurements are expensive and time-consuming. Traditional surrogate models often struggle to capture the complex interdependencies between different load measurements, especially when data is scarce or some measurements are unavailable. This paper addresses the challenge of creating robust flight load surrogate models by explicitly integrating known physics-based relationships (like those from flight mechanics) into the modeling process, aiming to improve prediction accuracy and data efficiency.

Methods

The core of this paper lies in extending Gaussian Process Regression (GPR) to a multi-output setting, specifically "joint multi-output GPR." While standard GPR models a single output, multi-output GPR can model several related outputs simultaneously. The key innovation here is the development of a framework to construct auto- and cross-covariance functions that are consistent with known physical laws. This allows the model to "learn" not just from the data but also from the underlying physics that govern the relationships between different flight load components.

The authors achieve this by:

  • Encoding Physics in Covariance Functions: Instead of treating outputs as independent, they define covariance functions that reflect known relationships (e.g., integral relationships between lift coefficient and spanwise force).
  • Handling Linear and Non-linear Relationships: They present methods for both linear and approximate methods for non-linear physics-based relationships between outputs.
  • Gradient-Enhanced Kriging (GEK) Extension: The work builds upon GEK, which incorporates gradient information, to a more general "integral enhanced kriging" and "any relationship between outputs" enhanced kriging.
  • Maximizing Marginal Likelihood: For optimal hyperparameter tuning, they maximize the marginal likelihood, balancing data fit with model complexity.

Datasets

The paper utilizes two types of datasets:

  • Simulated Data: Synthetic data is generated to test the methodology on quadratic relationships, where one output's data is intentionally hidden to evaluate the model's ability to predict missing information.
  • Flight Test Data: Real-world flight test data for a horizontal tail plane during a quasi-static longitudinal maneuver is used. This dataset includes measurements of the coefficient of lift ($C_z$) and spanwise lift ($k_z$), linked by a known physics-based integral relationship.

Key Insights and Ideas

  • Physics-Informed Surrogate Models are More Robust: Explicitly incorporating physics-based relationships significantly improves the predictive power of surrogate models, especially in scenarios with limited data or missing measurements.
  • Information Sharing is Key: Multi-output GPR effectively shares information between related outputs. When one output has data, it helps predict the other, even where direct measurements are absent.
  • Kernel Choice Matters: The choice of kernel (e.g., Squared Exponential vs. Matérn) significantly impacts the posterior predictions, highlighting the importance of considering the expected differentiability of the underlying physical processes.
  • Co-Kriging as a Special Case: The paper suggests that co-kriging can be viewed as a specific instance of their joint multi-output GPR framework, particularly when dealing with outputs of different fidelities.

Key Results

  • Improved Prediction Accuracy: On synthetic quadratic relationship data, joint multi-output GPR achieved a significant reduction in Root Mean Square Error (RMSE) for the output with hidden data (from 1.74 to 0.26 for $y_1$), demonstrating its ability to handle missing information.
  • Physically Consistent Models: The application to flight test data showed that the joint GPR model, incorporating flight mechanics knowledge, produced a more robust and physically consistent surrogate model for flight loads.
  • Handling Different Measurement Locations: The method can handle situations where measurements for different outputs are taken at different locations or times, as demonstrated by the flight test data analysis.

Limitations and Future Work

  • Scalability: The authors acknowledge that GPR, and by extension their multi-output extension, does not scale well to a very large number of input points.
  • Future Work: The paper suggests future research directions including investigating recent advances in scaling up GPR to handle more complex relationships and a larger number of inputs, potentially covering entire flight test datasets. They also envision using these robust surrogate models for applications like optimal sensor placement and Bayesian optimization.

Summary updated

Abstract

In this paper analytical methods to formally incorporate knowledge of physics-based equations between multiple outputs in a Gaussian Process (GP) model are presented. In Gaussian Processes a multi-output kernel is a covariance function over correlated outputs. Using a general framework for constructing auto- and cross-covariance functions that are consistent with the physical laws, physics-based relationships among several outputs can be imposed. Results of the proposed methodology for simulated data and measurement from flight tests are presented. The main contribution of this paper is the application and validation of our methodology on a dataset of flight tests, while imposing knowledge of flight mechanics into the model.
BibTeX
@misc{chiplunkar2016adding,
  author = {Ankit Chiplunkar and Emmanuel Rachelson and Michele Colombo and Joseph Morlier},
  title = {{Adding Flight Mechanics to Flight Loads Surrogate Model using Multi-Output Gaussian Processes}},
  year = {2016},
  doi = {10.2514/6.2016-4000},
}
Figure