Kolmogorov Arnold networks (KAN) for aerodynamic prediction: a comparison with MLPs and GNNs

2026-06-25 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Aerodynamic Modeling · Depth: Expert, quick

Summary

Kolmogorov Arnold networks (KANs), a deep neural network architecture adapting activation functions instead of affine transformation coefficients, were assessed for fluid dynamics surrogate modeling. Specifically, researchers compared KANs against deep multilayer perceptrons (MLPs) and Graph Neural Networks (GNNs) in predicting surface pressure distribution over subsonic and transonic airfoils. KAN models demonstrated good performance in predicting pressure coefficients and interpolating across Mach numbers and angles of attack. However, their performance was marginally inferior to a suitably trained MLP. GNNs achieved the best performance but required lengthier training. While optimal KAN models typically have lower complexity and faster training, they suffer from training instabilities and require proper hyperparameter optimization.

Key takeaway

For AI Scientists developing aerodynamic surrogate models, consider KANs for their lower complexity and faster training potential, but be prepared for significant hyperparameter optimization. While MLPs offer a robust, marginally superior baseline, and GNNs provide peak accuracy at higher training costs, KANs might suit scenarios prioritizing model interpretability or faster initial training cycles if stability issues are managed.

Key insights

KANs show promise for aerodynamic prediction but are marginally inferior to MLPs and less stable than GNNs.

Principles

• KANs adapt activation functions.
• MLPs offer strong baseline performance.
• GNNs achieve best accuracy.

Method

The study assessed KANs, MLPs, and GNNs for predicting surface pressure distribution over subsonic and transonic airfoils, comparing performance across Mach numbers and angles of attack.

In practice

• Evaluate KANs for fluid dynamics.
• Benchmark against MLPs and GNNs.
• Prioritize hyperparameter tuning for KANs.

Topics

• Kolmogorov Arnold Networks
• Aerodynamic Prediction
• Surrogate Modeling
• Multilayer Perceptrons
• Graph Neural Networks
• Fluid Dynamics

Best for: AI Scientist, Research Scientist

Related on AIssential

• Kolmogorov Arnold networks (KAN) for aerodynamic prediction: a comparison with MLPs and GNNs — KANs show promise in fluid dynamics but face stability issues and are marginally inferior to MLPs for aerodynamic prediction.
• A Practitioner's Guide to Kolmogorov-Arnold Networks — KANs offer enhanced expressivity and interpretability over MLPs by using learnable univariate basis functions on network edges.
• Ten Years of Trying to Learn from a Mesh That’s Too Large — Large CFD meshes pose representational challenges for surrogate models, requiring specialized architectures to capture localized physical features.
• Neural Network-Based Parametric Model Reduction for Predicting Turbulent Flow for Different Vehicle Geometries — A VAE-enhanced neural network model reduces computational cost for high-fidelity turbulent flow predictions around diverse vehicle geometries.
• A Geometry-Aware Triplane Field Network for Vehicle Aerodynamic Prediction — GTF-Net combines triplane representations with explicit geometric cues for accurate vehicle aerodynamic prediction.
• GAC-KAN: An Ultra-Lightweight GNSS Interference Classifier for GenAI-Powered Consumer Edge Devices — GAC-KAN offers an ultra-lightweight, high-accuracy GNSS interference classifier for GenAI edge devices via synthetic data and KANs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.