Enabling Efficient Crash Analyses in Aircraft Preliminary Design: An Explicit Non-Intrusive Methodology using Neural Differential Equations

Crash is a load case that is mandatory for the certification of aircrafts. To evaluate the influence of novel structural components, crash analyses should ideally be integrated into preliminary aircraft design, but the computational demands of detailed finite element simulations make this approach currently impractical. This paper presents a novel non-intrusive methodology that combines substructuring, mesh generalisation, proper orthogonal decomposition-based model order reduction, and neural differential equations to enable efficient crash analyses during early design stages. The methodology decomposes complex aircraft fuselage structures into manageable substructures, which are represented by individual surrogate models trained on finite element simulations. Mesh generalisation using basis functions decouples surrogate models from specific mesh configurations, enabling representation of diverse geometries within a unified framework. Proper orthogonal decomposition-based dimensional reduction compresses the solution space, whilst neural differential equations learn the underlying crash physics without explicit time parameterisation in the latent space. The methodology is demonstrated through a vertical strut substructure. The neural differential equation implementation successfully captures essential crash behaviour characteristics, though dynamic lag indicates areas for improvement in temporal response prediction. The framework provides a foundation for incorporating crashworthiness considerations into iterative preliminary design processes, potentially reducing reliance on costly late-stage design modifications.