A Multi-fidelity Double-Delta Wing Dataset and Empirical Scaling Laws for GNN-based Aerodynamic Field Surrogate

AI summary

Overview

Research area: Machine learning for aerospace engineering — specifically graph-neural-network (GNN) surrogates for aerodynamic field prediction, multi-fidelity dataset construction, and empirical scaling laws for scientific machine learning.

Technical level: Advanced. The paper assumes familiarity with computational fluid dynamics (CFD), vortex lattice methods, graph neural networks, and design-of-experiments theory, though the central question it asks is conceptually simple.

Scope (one sentence): The paper releases an open-source multi-fidelity double-delta wing aerodynamic dataset (2448 flow snapshots over 272 geometries) and uses nested subsets of it to measure how the prediction error of the MF-VortexNet GNN surrogate decays as training data and model size grow.

What This Paper Is About

Data-driven surrogate models can replace expensive aerodynamic simulations during aircraft design, but the field lacks both open multi-fidelity datasets and empirical guidance on how much training data is actually needed to reach a target accuracy. This paper addresses both gaps at once: it builds and releases a controlled, parametrically sampled double-delta wing dataset solved with a cheap vortex lattice method (VLM) and an expensive RANS CFD solver, then trains the MF-VortexNet surrogate on progressively larger subsets of that dataset to measure a data-scaling law. The goal is to let designers make informed decisions about how densely to sample a design space before committing to costly high-fidelity simulations.

Key Contributions

1. An open-source multi-fidelity double-delta wing aerodynamic dataset. 2448 flow snapshots across 272 unique geometries, evaluated at angles of attack from 11° to 19° (nine AOA values in 1° increments) at Ma = 0.3, using both a VLM solver and a RANS CFD solver. The release includes aerodynamic coefficients, surface solutions, volume solutions, and the CFD configuration files.

2. A nested Saltelli sampling scheme for the design of experiments. The geometry samples are generated from a Sobol sequence via SALib's sample.saltelli routine with skipvalue=16 and calc_second_order=False, producing a nested, extendable dataset that supports future variance-based sensitivity analysis and uncertainty quantification. A separate 16-geometry holdout test set is generated by Latin Hypercube Sampling.

3. An empirical data-scaling study of the MF-VortexNet surrogate. Six nested training datasets are constructed (spanning 40 to 1280 snapshots in the abstract's framing, drawn from 8 to 256 geometric configurations) and models with 0.1 to 2.4 million parameters are trained under a fixed training budget.

4. A power-law scaling law and a sampling-density heuristic. Test error decreases with data size with a power-law exponent of -0.6122, from which the authors estimate an optimal sampling density of roughly eight samples per dimension in a d-dimensional design space.

Main Findings

• Error decays as a power law in data size. Fitting the six nested training sets yields a test-error scaling exponent of -0.6122, which the authors describe as indicating efficient data utilization.

• A practical sampling-density rule of thumb. From the fitted law, the authors estimate that the optimal sampling density is approximately eight samples per dimension in a d-dimensional design space.

• Larger models appear to use data more efficiently. The results suggest improved data-utilization efficiency for larger surrogate models, implying a potential trade-off between dataset generation cost and model training budget.

• The dataset's design space is six-dimensional. The design variables are droop angle δ (-8° to 7°, nominal 0), inboard leading-edge sweep Λ_in (47° to 67°, nominal 65°), outboard leading-edge sweep Λ_out (35° to 60°, nominal 45°), outboard trailing-edge sweep Λ_te (0° to 25°, nominal 10°), break-chord fraction BW2 (0.32 to 0.47, nominal 0.4), and wing span B (1100 to 1300 inches, nominal 1200). Bounds were adapted from the SCALOS and HSCT programs and constrained to avoid non-physical geometries (e.g., negative tip chord), which lowers the Λ_in upper bound relative to SCALOS.

• Nested dataset levels. Level 1 through Level 6 correspond to base sequence lengths P = 1, 2, 4, 8, 16, 32, giving 8, 16, 32, 64, 128 and 256 geometry configurations, and dataset sizes of 72, 144, 288, 576, 1152 and 2304 snapshots respectively (each geometry evaluated at nine AOAs). Each level is a strict subset of the next.

• A mesh convergence study selected a 6.96 million cell mesh. Six surface-mesh refinement levels used maximum edge lengths of 0.1, 0.05, 0.0375, 0.025, 0.02 and 0.01 MAC, producing surface element counts of 1.01×10⁵, 1.05×10⁵, 1.08×10⁵, 1.18×10⁵, 1.25×10⁵ and 1.82×10⁵, and volume cell counts of 5.32, 6.13, 6.57, 6.96, 7.54 and 10.72 million. Area-averaged y+ values were 6.93, 3.50, 1.98, 1.66, 1.35 and 1.32. Level 4 (6.96×10⁶ volume cells) was within 1% of the finest grid for C_D and C_M and was applied to the full dataset.

• Total dataset cost. 310,835 CPU-hours produced the 2,448 snapshots, computed on a 32-core 2.4 GHz partition at the Stanford Research Computing Center (SRCC).

• Known fidelity limits of the solvers. The VLM (from SUAVE, based on Miranda et al.'s VORLAX, on a 30 × 32 chordwise-by-spanwise lattice) is accurate up to roughly 10° AOA; the SA-R RANS solver in SU2 degrades near a critical AOA of approximately 18° for some geometries, where vortex sheet oscillation or separated flow appear.

Methodology in Plain English

1. Define the shape. A double-delta (cranked delta) wing is described by six numbers: two leading-edge sweep angles, a trailing-edge sweep angle, where the inboard-to-outboard break chord sits, the span, and a "droop" angle that bends the leading-edge camber about the 30% chord point. The baseline airfoil comes from the RW23 wing root section. Inboard twist is a constant 1.5° leading-edge-up; outboard twist washes out linearly from 1.5° to -1.5°.

2. Choose which shapes to simulate. Rather than random or factorial sampling, the authors use a Sobol quasi-random sequence fed into Saltelli's scheme, which gives uniform design-space coverage, is reproducible, and — crucially — is nested: doubling the sample count adds new points without moving or invalidating the old ones. That nesting is what makes the scaling study possible without rerunning CFD. A separate 16-geometry holdout set is drawn by Latin Hypercube Sampling so the test set never disturbs the training set's uniform structure.

3. Solve each shape cheaply and expensively. Every geometry is run through the VLM solver (inviscid, compressibility-corrected, essentially free) and through the SU2 RANS solver with the Spalart-Allmaras model with rotation corrections (expensive, captures vortex and viscous effects). Meshes are generated automatically through OpenVSP's ComputeCFDMesh and Pointwise's T-Rex extrusion, with a trailing-edge source added for refinement.

4. Train a model to bridge the two fidelities. MF-VortexNet is a lightweight physics-informed GNN with a U-Net-like encoder-decoder built from multi-head graph-attention blocks. It takes the same lifting-surface lattice, represented as a graph whose nodes carry geometry, flow conditions, and low-fidelity pressure coefficients and whose edges encode panel-to-panel distances, and outputs corrected nodal pressures that approximate the CFD solution. Because it predicts a full field rather than a handful of scalars, any integrated quantity of interest can be post-processed afterward.

5. Measure how error shrinks with data. Six nested training sets and models of 0.1 to 2.4 million parameters are trained under a fixed budget, and the resulting test errors are fit to a power law.

6. Report the data cost. Alongside the dataset, the authors document the mesh convergence study, solver settings (JST scheme with coefficients 0.5 and 0.02, implicit Euler time advancement with adaptive CFL, FGMRES with ILU preconditioning and a residual tolerance of 10⁻⁷, continuity convergence at 10⁻⁶, and drag-coefficient changes below 10⁻⁵ over 100 Cauchy iterations), freestream conditions (285 K, static pressure 7.18×10⁴ Pa, gas constant 287.06, heat capacity ratio 1.4, Sutherland reference viscosity 1.715×10⁻⁵ Ns/m², laminar Prandtl 0.72 and turbulent Prandtl 0.9, Re_ref = 8.04×10⁷ based on a mean aerodynamic chord of 16 m), and a far-field boundary diameter of 40 MAC (630 m).

Why This Matters

Impact on research. Empirical scaling laws for scientific machine learning models remain rare, and physics-informed multi-fidelity GNN surrogates for field prediction have essentially none in the aerodynamic literature. A measured exponent of -0.6122 and a concrete sampling-density heuristic give the community something to plan experiments against instead of guessing, and the released dataset — with coefficients, surface and volume solutions, and CFD configuration files — provides a reproducible benchmark that the AIAA Applied Aerodynamics Surrogate Modeling discussion group datasets reportedly do not supply in a controlled parametric form.

Real-world applications.

• Conceptual and preliminary aircraft design, where designers need fast, arbitrary-quantity-of-interest feedback on configurations before committing to high-fidelity analysis.
• Aerodynamic shape optimization with stability constraints, particularly for low-speed, high-angle-of-attack configurations where vortex lift and vortex breakdown make linear methods unreliable.
• Supersonic transport and low-boom vehicle studies, since the design-space bounds were chosen by referencing the SCALOS and HSCT programs and nominal leading-edge sweeps of 68° and 45° at an aspect ratio of 2.5.
• Experiment planning for expensive CFD campaigns, where the sampling-density rule informs how many high-fidelity runs a team should budget before hitting diminishing returns.

Industry relevance. The multi-fidelity framing directly targets the new-product-introduction pipeline, where analyses move from crude conceptual tools to refined simulations and the gaps between fidelity levels need to be bridged. Because MF-VortexNet corrects a VLM field at a cost comparable to the VLM itself, it offers "quasi-CFD" pressure distributions at conceptual-design cost — and because it predicts fields rather than scalars, it can couple to solvers in other disciplines inside a multidisciplinary design optimization loop.

Future Directions

• Expand the dataset using the nested sampling structure. The Saltelli scheme was chosen precisely so that new samples can be added without rerunning existing experiments, and the authors explicitly frame the current work as a "preliminary" scaling study.
• Add higher-fidelity physics near the critical angle of attack. Flows near approximately 18° AOA involve massive separation and unsteady phenomena that the SA-R RANS model cannot reliably capture; detached-eddy simulation is named as a promising but presently prohibitive extension.
• Test whether the -0.6122 exponent and the eight-samples-per-dimension rule generalize beyond this six-dimensional double-delta wing design space, this surrogate architecture, and this pair of fidelity levels.
• Sharpen the trade-off between data cost and model capacity. The observation that larger models use data more efficiently raises an open question about how to allocate a fixed budget between generating CFD snapshots and training larger surrogates.

Target Audience

Aerospace engineers and researchers working on surrogate modeling, multidisciplinary design optimization, and conceptual aircraft design; machine learning researchers interested in scaling laws for physics-informed and graph-based models; and practitioners who need to justify CFD dataset budgets or who want a ready-made multi-fidelity benchmark with paired VLM and RANS solutions for a vortex-dominated configuration. Readers without a background in CFD or graph neural networks will find the dataset description accessible but the scaling analysis demanding.

Note on the source text: the provided paper content is truncated during the discussion of the CFD results (Section 3.4.3, describing Figure 6), so any later results, discussion, and conclusion sections are not reflected above. Figures 6 and any subsequent tables are likewise not summarized, and specific numerical error values from the scaling study beyond the exponent and the sampling-density estimate are not reported in the available text.

Authors’ abstract

Data-driven surrogate models are increasingly adopted to accelerate vehicle design. However, open-source multi-fidelity datasets and empirical guidelines linking dataset size to model performance remain limited. This study investigates the relationship between training data size and prediction accuracy for a graph neural network (GNN) based surrogate model for aerodynamic field prediction. We release an open-source, multi-fidelity aerodynamic dataset for double-delta wings, comprising 2448 flow snapshots across 272 geometries evaluated at angles of attack from 11 (degree) to 19 (degree) at Ma=0.3 using both Vortex Lattice Method (VLM) and Reynolds-Averaged Navier-Stokes (RANS) solvers. The geometries are generated using a nested Saltelli sampling scheme to support future dataset expansion and variance-based sensitivity analysis. Using this dataset, we conduct a preliminary empirical scaling study of the MF-VortexNet surrogate by constructing six training datasets with sizes ranging from 40 to 1280 snapshots and training models with 0.1 to 2.4 million parameters under a fixed training budget. We find that the test error decreases with data size with a power-law exponent of -0.6122, indicating efficient data utilization. Based on this scaling law, we estimate that the optimal sampling density is approximately eight samples per dimension in a d-dimensional design space. The results also suggest improved data utilization efficiency for larger surrogate models, implying a potential trade-off between dataset generation cost and model training budget.