HiLiftAeroML: Scalable High-Lift Surrogates

• HiLiftAeroML is a comprehensive ecosystem for high-fidelity surrogate modeling and aerodynamic analysis, integrating large-scale CFD datasets with advanced surrogate architectures.
• It employs diverse techniques such as MLPs, Gaussian processes, graph neural networks, and neural operators to achieve scalable, accurate predictions for high-lift configurations.
• The platform features interactive, language-driven tools that enable real-time data retrieval, uncertainty quantification, and aerodynamic design optimization.

HiLiftAeroML is a comprehensive ecosystem for high-fidelity surrogate modeling and interactive analysis of high-lift aircraft aerodynamics, integrating a publicly available large-scale CFD dataset, advanced surrogate architectures, and a retrieval-augmented language-driven user interface. The system addresses the critical need for scalable, accurate, and interpretable prediction and design tools in the high-lift regime, where standard RANS-based approaches lack fidelity, and existing datasets or surrogate models are limited in coverage, accessibility, or resolution.

1. Dataset Composition and Generation

The core of HiLiftAeroML is a high-fidelity CFD dataset comprising 1,800–2,550 cases of high-lift wing-body geometries based on the NASA Common Research Model (CRM) high-lift configuration (Ashton et al., 19 May 2026). Geometric variations span eight independent parameters, including inboard/outboard slat and flap deflections and gap multipliers, sampled via Latin Hypercube in an 8D design space. Each geometry is simulated at 10 angles of attack, from 4° through 22°, capturing pre-stall, near-stall, and post-stall conditions.

All field data are generated using a GPU-accelerated, wall-modeled large-eddy simulation (WMLES) with the Fidelity Charles solver and solution-adapted Voronoi-based meshes refined to 300–500 million cells per case. Key simulation features include:

• Dynamic Smagorinsky SGS model for subgrid stresses.
• Equilibrium wall modeling for high-resolution near-wall treatment.
• Domain is a half-span aircraft in a hemispherical volume, with plug-flow inlet, non-reflecting outlet, and free-stress symmetry conditions.
• Adaptivity is driven via error indicators (strain rate, Reynolds-stress residual) for statistical convergence of CL, CD, CM.

Each exeperiment folder contains geometry files (.stp, .stl), time-averaged volume and surface fields, integrated force/moment CSVs, geometry/condition metadata, wall-shear stress visualizations, and convergence diagnostics. The full dataset is open-source under CC-BY-4.0 and accessible via HuggingFace, with tooling provided for download and data partitioning.

2. Surrogate Modeling Methodologies

HiLiftAeroML supports several classes of surrogate architectures leveraging the available rich geometric and field data (Ashton et al., 19 May 2026, Giral et al., 7 May 2026):

• End-to-End Prediction: Multi-layer perceptrons (MLPs) mapping [geometry, AoA] vectors directly to integrated aerodynamic coefficients (CL, CD, CM).
• Gaussian Process Regression: For low-cardinality design subsets, direct regression of global coefficients or local pressure distributions.
• Graph Neural Networks: Applied on surface mesh graphs for field distributions, especially pressure.
• Neural Operators: Such as DeepONet and Fourier Neural Operator, trained for volumetric field mapping.
• Transformers and Point Clouds: Methods like GeoTransolver and AeroJEPA operate on unstructured point clouds, scaling to millions of points per case with attention-based encoding and decoding.

AeroJEPA, in particular, demonstrates scalable field-level surrogate accuracy at high resolutions (12–15M surface points per sample), with a joint-embedding architecture decoupling geometric encoding from explicit field reconstruction (Giral et al., 7 May 2026).

Recommended pre-processing steps include mesh coarsening, geometric normalization, freestream standardization of flow variables, and dimensionality reduction via PCA, POD, or autoencoders for latent representation learning.

3. Joint-Embedding Architectures and Latent Space Organization

AeroJEPA ("Joint-Embedding Predictive Architecture") exemplifies a state-of-the-art approach for semantic and scalable surrogate modeling using HiLiftAeroML (Giral et al., 7 May 2026). Its formulation:

• Separates the encoding of geometry/context (point cloud $\mathcal{P}$) and field/target (subsampled CFD field $\mathcal{F}$) into sets of context and target tokens via Point Transformer backbones.
• Predicts flow latents conditioned on context tokens and operating conditions, with an optional continuous implicit neural representation (INR) decoder reconstructing field values at arbitrary query points.
• Trains using a combination of latent-matching loss, field reconstruction loss, and the SIGReg regularizer to prevent token collapse.
• Achieves a mean surface pressure RMSE of 0.0046 (0.48% rel. L₂), which is four times lower than leading baselines (e.g., FigConvUNet, GeoTransolver), and reduces inference cost by approximately 35% compared to the fastest alternative surrogates.

Latent space analyses demonstrate that context and predicted latents capture geometric and aerodynamic variations, allowing for linear probing of control surface settings and aerodynamic coefficients ($R^2$ up to 0.996 for CD), as well as concept-vector operations for controlled design perturbation and smooth field interpolation between configurations.

4. Model Evaluation, Benchmarks, and Practical Use

HiLiftAeroML surrogates are benchmarked under in-distribution and out-of-distribution regimes with standardized train/validation/test splits by geometry, angle of attack, and design parameters (Ashton et al., 19 May 2026, Giral et al., 7 May 2026). Use cases range from rapid evaluation of lift/drag curves to high-resolution surface/volumetric reconstructions.

Model	Surface p RMSE (max GT)	MAE (max GT)	Inference Cost (TFLOPs)	Key Features
FigConvUNet	0.0197	0.0120	88.3	Chunked field U-net
GeoTransolver	0.0277	0.0184	309.1	Point-cloud transformer
AeroJEPA	0.0046	0.0023	57.0	Joint-embedding, tokens

AeroJEPA enables query-time field evaluation at any point or surface location, with the latent space supporting physically interpretable and differentiable surrogate surfaces—facilitating design analysis, gradient-based optimization, and parametric studies beyond the resolution or density available in tabulated CFD.

5. Interactive Interfaces and Retrieval-Augmented Generation

The HiLiftAeroML platform integrates a domain-specialized LLM ("HiLiftAeroGPT"), the ASPIRE experimental database of airfoil pressure distributions, and the ADAPT deep-kernel learning engine (Lee et al., 2024). The system enables:

• Natural-language question parsing and intent classification (retrieve, predict, integrate).
• Retrieval-augmented generation (RAG) workflows, searching over ASPIRE metadata (~2,800 measured pressure distributions covering 68 airfoils, $\alpha\in[-30°,30°]$, $M_\infty\in[0,1.0]$, $Re_c\sim10^6$).
• On-demand pressure and coefficient prediction with uncertainty quantification and Monte Carlo propagation, handled by the ADAPT model (deep-kernel GP with neural latent embedding).
• Automatic integration of pressure for $c_\ell, c_d, c_m$ with $\mathrm{MAE}_{\rm enclosed} = 0.029$ for $C_p$ fields on test airfoils.
• Output formatting with LaTeX-formatted equations, tables, and plots, supporting API and chat-based workflows.

Example Q&A sessions document the workflow of data retrieval, prediction, integration, and presentation—including uncertainty bandwidth and historical dataset comparisons.

6. Access, Licensing, and Replication

HiLiftAeroML provides open access to data and code. The dataset is released under the Creative Commons Attribution 4.0 International License (CC-BY-4.0), permitting commercial and noncommercial use with attribution (Ashton et al., 19 May 2026). The repository includes scripts for batch download, data partitioning, and standardization. Surrogate modeling pipelines, training protocols (e.g., AdamW, learning-rate schedules), and full model definitions for AeroJEPA and baseline architectures are available to enable exact replication (Giral et al., 7 May 2026).

Replication tips emphasize efficient point-cloud subsampling (farthest point sampling), message-passing tokenization, loss computation, and hardware requirements (NVIDIA H200 class GPU, 48 hr training per JEPA model). Diagnostics recommended are $\mathcal{L}_{\mathrm{lat}}$, $\mathcal{F}$0, SIGReg singular-value spread, and periodic surrogate error on held-out geometries.

7. Scientific Impact and Future Directions

HiLiftAeroML constitutes the first fully open high-fidelity high-lift aircraft CFD dataset paired with modern surrogate modeling and LLM-driven analysis workflows. It addresses a central challenge in aerospace design—the need for accurate, scalable surrogates in the nonlinear, transitory physics regime of multi-element high-lift wings. The latent organization, interpretability, and scalability demonstrated by models such as AeroJEPA suggest new research avenues in differentiable design, latent-space optimization, and uncertainty-aware surrogate integration. A plausible implication is the emergence of interactive, language-driven design assistants for aerospace engineers, able to perform data-driven and physically consistent aerodynamic analyses at full 3D field resolution (Lee et al., 2024, Giral et al., 7 May 2026, Ashton et al., 19 May 2026).