Dataset for the paper "Multi-fidelity transonic aerodynamic loads estimation using Bayesian neural networks with transfer learning"

This repository provides the multi-fidelity aerodynamic datasets used in:A. Vaiuso, G. Immordino, M. Righi, A. Da Ronch. "Multi-fidelity transonic aerodynamic loads estimation using Bayesian neural networks with transfer learning." Aerospace Science and Technology, 163 (2025) 110301.https://doi.org/10.1016/j.ast.2025.110301

The datasets enable training, validation, and benchmarking of a surrogate model for transonic aerodynamic load prediction with quantified uncertainty. They include low-, mid-, and high-fidelity aerodynamic data for a finite wing test case in transonic conditions and a full eVTOL aircraft configuration.

1. Finite Wing – Benchmark Supercritical Wing (BSCW)

Design space: Mach number ∈ [0.70 , 0.84] and Angle of Attach AoA ∈ [0, 4] deg.

Low-fidelity (LF): 625 samples generated with XFoil (panel method with 3D corrections).

Mid-fidelity (MF): 49 RANS simulations with SU2 (coarse 2.5M-cell hybrid grid, Spalart–Allmaras turbulence model). Grid Convergence Index ≈ 5.4%.

High-fidelity (HF): 58 RANS simulations with SU2 (fine 15.6M-cell grid, GCI ≈ 0.6%). Of these, 7 samples were used for model fine-tuning, while 51 were retained as an independent HF test set.

Outputs: lift coefficient (CL) and pitching moment coefficient (CM) relative to 30% chord.

2. Full–Configuration eVTOL Vehicle

Geometry: 5-propeller configuration with four symmetric lift propellers (front/rear) and one tail-mounted pusher.

Design space: AoA ∈ [−180 , 180] deg, freestream velocity U ∈ [0 , 40] m/s, and propeller RPM up to ±4000.

Low-fidelity (LF): 2000 samples from Blade Element Momentum (BEM) methods with momentum-theory-based induced velocity corrections. 

Mid-fidelity (MF): 250 simulations using the DUST framework (Vortex Particle Method with nonlinear lifting-line). Between 30–40 lifting lines per blade; wing discretised with ~3000 panels per side.

Outputs: thrust (KT) and torque (KQ) coefficients for each of the five propellers.

PurposeThese datasets support the development of multi-fidelity machine-learning frameworks for transonic aerodynamics, providing training material for uncertainty-aware surrogate models. They are also relevant for applications in aeroelasticity, flight dynamics, and design of unconventional aircraft (e.g. eVTOLs).

Contents

Raw aerodynamic datasets at LF, MF, HF fidelities.

Partitioning into training, validation, and independent test sets as described in the paper.

Associated aerodynamic coefficients across Mach–AoA (BSCW) and AoA–velocity–RPM (eVTOL) parameter spaces.

KeywordsMulti-fidelity modelling; Bayesian neural networks; transfer learning; uncertainty quantification; Benchmark Supercritical Wing; eVTOL aerodynamics; CFD; reduced-order modelling.