Physics-informed machine learning for near real-time stress prediction on a structural component: application for landing gears

Abstract

Lightweight design constitutes a pivotal research and development objective for next-generation landing gear systems. Nevertheless, achieving reduced weight while maintaining structural safety and reliability presents considerable challenges. The establishment of a digital twin (DT) for structural health monitoring (SHM) offers a promising approach to address these concerns across the design, testing, and operational lifecycle of landing gears. In this study, we develop a physics-informed neural network (PINN) model for near real-time stress prediction on the drag strut of a nose landing gear (NLG), specifically for an A320-type aircraft, serving as a foundational component of a DT system. The proposed PINN framework directly outputs displacement fields while deriving stresses as secondary quantities, effectively incorporating the fundamental equations of linear elasticity into the loss function. Displacement boundary conditions, informed by finite element method (FEM) simulations, are integrated as penalty terms to enhance trainability and physical consistency. The training dataset is constructed using load cases statistically representative of actual landing gear operations, with high-fidelity FEM providing corresponding displacement and stress references. The model demonstrates strong predictive accuracy, with relative errors between 5% and 7% compared to FEM results, and significantly outperforms both pure stress-output PINNs and conventional deep neural networks (DNNs). Moreover, the trained PINN achieves inference times within seconds under time-varying loads, highlighting its capability for near real-time stress monitoring. This work underscores the potential of physics-informed machine learning for enhancing DT-enabled SHM systems in safety-critical aerospace structures.