HiLiftAeroML: High-Fidelity Computational Fluid Dynamics Dataset for High-Lift Aircraft Aerodynamics

This paper introduces HiLiftAeroML, the first open-source high-fidelity CFD dataset featuring 1,800 GPU-accelerated LES simulations of NASA's CRM high-lift geometry, designed to accelerate the development of AI surrogate models for aerospace applications.

The Gist

Imagine you are trying to teach a robot how to fly a complex airplane. To do this, you need to show it thousands of examples of how the air moves around the wings, especially when the plane is taking off or landing. These moments are tricky because the air gets messy, swirling, and separating from the wings in chaotic ways.

For a long time, scientists have used two main ways to study this:

1. Wind Tunnels: Building physical models and blowing real air at them. This is accurate but incredibly expensive and slow.
2. Computer Simulations (CFD): Using math to predict the air. The standard method is fast but often gets the messy parts wrong, like a blurry photo. A better method exists that takes high-definition photos of the air, but it usually takes supercomputers weeks to generate just one picture.

The Problem: To train a smart AI (a "surrogate model") to predict these messy airflows instantly, you need a massive library of these high-definition pictures. But until now, that library didn't exist for complex airplanes.

The Solution: HiLiftAeroML
This paper introduces HiLiftAeroML, a massive, free, open-source library of 1,800 high-definition "snapshots" of air flowing around a specific type of airplane (the NASA Common Research Model).

Here is how they built it, using some simple analogies:

1. The Airplane: A Shape-Shifting Lego Set

The researchers didn't just use one airplane. They used a digital version of the NASA "Common Research Model" (CRM), which is like a standard Lego airplane used by scientists worldwide.

• The Twist: They made the Lego pieces move. They created 180 different versions of this airplane by changing the angles of the flaps and slats (the little wings on the front and back that pop out during takeoff and landing).
• The Weather: For each of those 180 shapes, they simulated the air hitting the plane at 10 different angles (from a gentle approach to a steep climb).
• The Result: 1,800 unique scenarios (180 shapes × 10 angles).

2. The Camera: A Super-Sharp Lens

Most computer simulations use a "blurry" lens (called RANS) that averages out the chaos. It's like watching a sports game through a foggy window; you see the players moving, but you miss the individual spins and collisions.

For this dataset, the authors used a Wall-Modeled Large Eddy Simulation (WMLES).

• The Analogy: Think of this as a 4K, slow-motion camera that captures every single swirl and eddy of the air.
• The Cost: This "camera" is so powerful that it requires a grid of 300 to 500 million tiny cells (pixels) just to cover the airplane. To put that in perspective, a standard simulation might use 10 million cells. This is like upgrading from a standard-definition TV to a massive, ultra-high-definition screen.
• The Hardware: They ran these simulations on NVIDIA GPUs (the same powerful chips used for gaming and AI), which acted like a fleet of super-fast cameras snapping these pictures.

3. The Library: Free for Everyone

The authors didn't keep these 1,800 high-definition snapshots to themselves. They put the entire library on the internet (HuggingFace) for anyone to download for free.

• What's inside: You get the 3D shape of the airplane, the "blurry" average forces (lift and drag), and the detailed "high-def" data of the air pressure and speed inside and around the plane.
• The Goal: They want AI researchers to use this library to train their own "flight robots." Once an AI learns from these 1,800 perfect examples, it should be able to predict how the air behaves on new airplane designs in a split second, without needing to run the expensive, slow simulation again.

4. Did it Work? (The Quality Check)

Before releasing the library, the authors checked their work against real-world wind tunnel experiments.

• The Test: They compared their computer "photos" of a specific landing configuration against actual photos taken in a wind tunnel.
• The Result: Their high-definition simulation matched the real-world data very well, especially for the tricky parts like "drag" (air resistance) and "pitching moment" (how the nose wants to tilt). This proves their "camera" was sharp enough to capture the real physics.

Summary

In short, the authors built the first-ever "high-definition" library of airplane aerodynamics for takeoff and landing scenarios. They used the most advanced, expensive, and accurate computer methods available to generate 1,800 examples. By making this data free, they hope to help engineers and AI developers build smarter, faster tools to design safer and more efficient airplanes in the future.

What the paper does NOT claim:

• It does not claim that AI has already replaced wind tunnels (it's a tool to help, not a replacement yet).
• It does not claim to have solved the physics of every possible airplane (it focuses on this specific NASA model).
• It does not claim to have simulated full-scale flight conditions (the data is based on wind-tunnel scale conditions).

Technical Summary

Technical Summary: HiLiftAeroML

Problem Statement
The optimization of high-lift aircraft systems is critical for safety and operational performance, particularly during take-off and landing where complex, three-dimensional unsteady flows involving massive separation and vortex dynamics occur. Historically, the aerospace industry has relied on Reynolds-Averaged Navier-Stokes (RANS) CFD solvers due to their low computational cost; however, RANS frequently fails to resolve the physics inherent to high-lift configurations, necessitating continued reliance on expensive physical wind-tunnel testing. While high-fidelity scale-resolving methods like Wall-Modeled Large Eddy Simulation (WMLES) offer superior accuracy, their computational burden (often requiring an order of magnitude more resources than RANS) makes them prohibitively expensive for the thousands of simulations required for design exploration.

Concurrently, the field of Artificial Intelligence (AI) is increasingly turning toward surrogate models to accelerate design cycles. However, the development of robust AI surrogates for aerospace is hindered by a lack of open-source, high-fidelity training data. Existing public datasets are almost exclusively limited to RANS simulations of cruise conditions or simplified geometries, failing to address the high-lift regime where complex separation physics dominate.

Methodology
To address this data gap, the authors generated the HiLiftAeroML dataset, comprising 1,800 high-fidelity CFD samples. The methodology involved the following key components:

• Geometry: The dataset is based on the NASA Common Research Model High-Lift (CRM-HL) geometry. Unlike previous workshops that used specific wind-tunnel models, this work utilized a theoretical reference geometry with parametrically defined mechanical features (slat brackets, fairings) to decouple the simulation from physical hardware constraints.
• Parametric Variation: The dataset covers 180 unique geometry variants generated via Latin Hypercube Sampling. These variants manipulate eight independent parameters: inboard/outboard slat deflection and gap, and inboard/outboard flap deflection and gap. Each geometry was simulated at 10 Angles of Attack (AoA) ranging from 4° to 22° in 2° increments to capture pre-stall and post-stall regimes.
• Solver and Physics: Simulations were performed using the Fidelity Charles flow solver, an explicit, unstructured, finite-volume solver for the compressible Navier-Stokes equations. The study employed a Wall-Modeled Large Eddy Simulation (WMLES) approach with a dynamic Smagorinsky subgrid-scale model and an equilibrium wall model. This approach resolves large-scale eddies while modeling subgrid interactions, offering superior fidelity over RANS for separated flows.
• Mesh and Adaptation: The domain was discretized using Fidelity Stitch, a Voronoi-based unstructured mesh generator. A solution-adaptation algorithm was employed to refine the grid based on flow features, resulting in meshes containing between 300 million and 500 million control volumes.
• Data Processing: Time-averaging was performed after flushing initial transients. Simulations for AoA ≤ 12° ran for 15 Convective Time Units (CTUs), while higher AoA cases (≥ 12°) ran for 30–200 CTUs to ensure statistical convergence. The raw volumetric data was exported in an Octree format to manage file sizes while preserving accuracy.

Key Contributions

1. First Open-Source High-Fidelity High-Lift Dataset: HiLiftAeroML is the first open-source dataset dedicated to high-lift aerodynamics using full-aircraft WMLES simulations. It provides 1,800 samples including geometries, time-averaged volume and surface variables, and integral forces.
2. High-Fidelity Resolution: The dataset utilizes solution-adapted grids (300M–500M cells) and WMLES, ensuring the highest possible accuracy for the flight envelope portions where RANS is known to struggle.
3. Comprehensive Parametric Space: The dataset covers a wide range of geometric perturbations (slat/flap deflections and gaps) and flow conditions (AoA 4°–22°), explicitly including deep-stall and massive separation regimes.
4. Benchmarking Framework: The authors provide deterministic train, validation, and test splits designed to evaluate specific machine learning capabilities, including data efficiency, generalization to unseen geometries, and out-of-distribution (OOD) extrapolation across angles of attack and deflection settings.

Results and Validation
The dataset was validated against experimental wind-tunnel data (Mouton et al., 2024) for the LDG-HV configuration.

• Force Prediction: The WMLES results on the adapted grid showed reasonable agreement with experimental lift, drag, and pitching moment coefficients. Notably, grid adaptation led to significant improvements in drag and pitching moment predictions compared to the baseline grid.
• Flow Physics: Qualitative comparisons of near-surface flow patterns (e.g., at AoA = 18°) demonstrated that the simulations correctly captured outboard wedge-shaped separation patterns and flow separation on flaps, matching experimental oil-film visualizations.
• Dataset Diversity: Analysis of the dataset revealed that while lift coefficients at low AoA are primarily driven by flap deflection, the sensitivity shifts to slat deflection at high AoA (near stall), reflecting the complex, non-linear aerodynamic behavior of the high-lift regime.

Significance and Claims
The paper positions HiLiftAeroML as a critical resource to accelerate the research and development of AI surrogate models within the aerospace industry. By providing access to "hard-to-simulate" high-lift regimes with high-fidelity data, the authors aim to:

• Enable the training of ML architectures on industrially relevant, unsteady flow physics that are currently under-explored.
• Allow engineers to rapidly filter thousands of designs, isolating promising candidates for rigorous validation.
• Serve as a benchmark for the 6th AIAA High-Lift Prediction Workshop and the broader community to assess ML approaches against high-fidelity CFD.

The authors maintain a modest tone regarding the dataset's limitations, acknowledging that the simulations are at wind-tunnel scale (Reynolds number $1.6 \times 10^6$) rather than full-flight scale, and that the flow is modeled as fully turbulent without explicit laminar-turbulent transition. Nevertheless, the dataset represents a significant step toward bridging the gap between high-fidelity physics and rapid AI-driven design cycles.