Adaptive-Quadratic-Neural-Network-Based Multifidelity Modeling Approach for Buckling of Stiffened Panels

What is it about?

This study focuses on predicting the buckling load of stiffened panels, which is a key structural design problem in aerospace engineering. Instead of relying only on expensive high-fidelity finite element simulations, the paper combines high- and low-fidelity simulation data in a multifidelity framework and trains neural-network surrogate models to estimate buckling loads more efficiently. The main contribution is the use of quadratic neural networks with adaptive activation functions, which improve prediction quality and learning efficiency compared with conventional neural-network approaches. The study evaluates several multifidelity correction strategies and shows that the best ratio-based model can achieve high accuracy at much lower computational cost than a purely high-fidelity workflow.

Featured Image

Why is it important?

Buckling analysis is central to the design of lightweight aerospace structures, but repeated high-fidelity finite element analysis is too costly for large design-space exploration and optimization. This paper shows that multifidelity surrogate modeling can reduce that burden substantially while preserving strong predictive performance. More specifically, the work shows that quadratic neural networks combined with adaptive activation functions improve convergence speed, reduce model complexity, and deliver more accurate multifidelity predictions than standard multilayer perceptrons. That makes the approach useful for faster structural screening, optimization, and design iteration in engineering settings where many candidate panel configurations must be evaluated.

Perspectives