A Kernel-based Resource-efficient Neural Surrogate for Multi-fidelity Prediction of Aerodynamic Field

Surrogate models provide fast alternatives to costly aerodynamic simulations and are extremely useful in design and optimization applications. This study proposes the use of a recent kernel-based neural surrogate, KHRONOS. In this work, we blend sparse high-fidelity (HF) data with low-fidelity (LF) information to predict aerodynamic fields under varying constraints in computational resources. Unlike traditional approaches, KHRONOS is built upon variational principles, interpolation theory, and tensor decomposition. These elements provide a mathematical basis for heavy pruning compared to dense neural networks. Using the AirfRANS dataset as a high-fidelity benchmark and NeuralFoil to generate low-fidelity counterparts, this work compares the performance of KHRONOS with three contemporary model architectures: a multilayer perceptron (MLP), a graph neural network (GNN), and a physics-informed neural network (PINN). We consider varying levels of high-fidelity data availability (0%, 10%, and 30%) and increasingly complex geometry parameterizations. These are used to predict the surface pressure coefficient distribution over the airfoil. Results indicate that, whilst all models eventually achieve comparable predictive accuracy, KHRONOS excels in resource-constrained conditions. In this domain, KHRONOS consistently requires orders of magnitude fewer trainable parameters and delivers much faster training and inference than contemporary dense neural networks at comparable accuracy. These findings highlight the potential of KHRONOS and similar architectures to balance accuracy and efficiency in multi-fidelity aerodynamic field prediction.

💡 Research Summary

The paper introduces KHRONOS, a kernel‑based neural surrogate designed for multi‑fidelity prediction of aerodynamic surface pressure coefficient fields on 2‑D airfoils. The authors address the challenge that high‑fidelity (HF) CFD simulations (AirfRANS) are expensive, while low‑fidelity (LF) models (NeuralFoil) are cheap but often inconsistent with HF data, especially in regions of stall or strong separation. Traditional multi‑fidelity approaches either rely on co‑kriging, latent‑space alignment, or deep architectures that require large numbers of trainable parameters and substantial computational resources.

KHRONOS departs from these methods by representing the high‑dimensional input‑output mapping as a tensor product of low‑dimensional subspaces. Each subspace is approximated by a kernel interpolation function whose parameters (grid points k and per‑dimension scaling γ_i) are learned via variational principles. This formulation yields a mathematically grounded “pruned” network: the number of trainable parameters grows linearly with the number of input dimensions rather than quadratically or exponentially as in dense MLPs or GNNs. The kernel expansion is auto‑differentiable, enabling gradient‑based training while preserving the physical smoothness inherent in variational formulations.

For multi‑fidelity learning, the LF dataset is generated by evaluating NeuralFoil on the same airfoil geometries and flow conditions used in the HF CFD runs. The LF fields are interpolated onto the HF mesh to create a consistent surface representation. KHRONOS then learns a residual surrogate x_Δ that directly maps the LF prediction to the HF target, avoiding the need for an explicit latent space that aligns disparate meshes or turbulence models. This residual learning is performed on a small set of HF samples (0 %, 10 %, 30 % of the total 1,000 CFD cases) while the majority of the training data comes from the LF model.

The experimental protocol varies two axes: (1) the proportion of HF data available, and (2) the complexity of the geometry parameterization (5, 10, 15 control points). Ten‑fold cross‑validation is used for each configuration, and three baseline architectures are compared: a multilayer perceptron (MLP), a graph neural network (GNN) that operates on the surface mesh, and a physics‑informed neural network (PINN) that incorporates the incompressible RANS equations as soft constraints. All models are trained on the same pre‑processed data and hyper‑parameter search space to ensure a fair comparison.

Results show that KHRONOS consistently attains a coefficient of determination R² ≥ 0.80 across all test cases, reaching R² ≈ 0.90 in the most challenging geometry setting. In contrast, the baselines require substantially more trainable parameters (approximately 20–30 × more) to achieve comparable R² values. KHRONOS reduces the parameter count by 94 %–98 % relative to the baselines, leading to dramatically faster training (3–15 seconds per fold) and inference (2.44–3.64 ms per prediction). The mean absolute error (MAE) is also reduced by 5 %–12 % compared with the best baseline.

A particularly insightful part of the study focuses on 265 “hard” cases where the LF model alone yields R² < 0.7 due to complex geometry or flow conditions near stall. Even with only a handful of HF samples, KHRONOS’s residual learning recovers most of the lost accuracy, demonstrating robustness to LF‑HF inconsistency.

The authors discuss the theoretical advantages of combining variational calculus, kernel interpolation, and tensor decomposition: (a) the variational foundation guarantees an energy‑optimal solution, (b) kernel interpolation provides smooth, physics‑consistent reconstructions without dense parameterization, and (c) tensor decomposition enables separable treatment of each input dimension, which is crucial for scalability. They also acknowledge limitations: the current work is restricted to 2‑D airfoil data, and extending to 3‑D configurations will require careful management of kernel grid dimensionality and possible hierarchical tensor schemes.

In conclusion, KHRONOS offers a compelling solution for resource‑constrained aerodynamic surrogate modeling. By leveraging a mathematically principled kernel expansion and a residual multi‑fidelity strategy, it achieves high accuracy with a fraction of the computational cost of state‑of‑the‑art dense neural networks. This makes it especially attractive for design optimization, uncertainty quantification, and real‑time control applications where rapid, reliable field predictions are essential.