Abstract
In this contribution, a 1-D surrogate model of variable fidelity based on Gaussian process regression, applied to the aerodynamic analysis within the context of conceptual aircraft design is presented. For the low-fidelity aerodynamic analysis semi-empirical methods combined with a lifting line method are used. For the higher-fidelity analysis a RANS solver is used. Furthermore an adaptive sampling strategy based on prior knowledge from the aircraft design process is elaborated. The aerodynamic analysis is applied to the CeRAS short-range aircraft CSR-02. The impact of the enhanced aerodynamic analysis is investigated by simulated off-design missions comparing the estimated fuel burn with the low-fidelity aerodynamic analysis.
Introduction
Conceptual aircraft design requires evaluating thousands of configurations rapidly. Traditionally, this process relies on "Level 0-1" methods—semi-empirical formulas that are fast but often inaccurate for unconventional designs. Conversely, high-fidelity "Level 3" methods like Reynolds-Averaged Navier-Stokes (RANS) CFD offer precision but are computationally prohibitive for iterative loops. This research introduces a variable-fidelity surrogate model that bridges this gap, using Gaussian Process (GP) regression to correct fast, low-fidelity models with a minimal number of strategic high-fidelity data points.
Methodology: Variable-Fidelity and Mission-Informed Sampling
The core of this approach is the integration of low-fidelity (LF) data as a "prior mean" for a Gaussian Process. Instead of training from scratch, the GP learns the discrepancy between the LF model and high-fidelity (HF) RANS data.
A key innovation is the mission-informed adaptive sampling strategy. Rather than sampling the design space uniformly, the authors utilize a "bias function" derived from simulated flight missions. By identifying the lift coefficients ($C_L$) where the aircraft operates most frequently (e.g., during cruise), the model prioritizes expensive CFD simulations at these critical points. This ensures that uncertainty is minimized where it has the greatest impact on performance metrics like fuel burn.
Case Study and Results
The method was validated using the CeRAS CSR-02, a short-range reference aircraft. Low-fidelity polars were generated via UNICADO’s calculatePolar tool, while high-fidelity data was produced using the DLR-TAU RANS solver.
As shown in Figure 4, the surrogate model acts as a localized "nudge," correcting the empirical baseline toward the CFD truth. Even with a single HF sample at the cruise point, the model's uncertainty was reduced by approximately 26 drag counts. The study demonstrated that the surrogate model converges rapidly; by the fifth HF sample, uncertainty in the region of interest was nearly eliminated. Crucially, comparing the enhanced surrogate to the standard LF model revealed a 3% difference in estimated fuel burn for a 2500 nm mission. In conceptual design, a 3% discrepancy is significant and could fundamentally alter the perceived viability of an aircraft project.
Limitations and Future Work
Currently, the model is 1-D, focusing on the relationship between lift and drag. Future iterations will expand to multi-dimensional surrogates incorporating Mach number, altitude, and complex geometries including nacelles and pylons. Additionally, the sampling strategy currently assumes a fixed mission profile; future work may explore adaptive sampling for varying mission requirements.
Summary updated