Stop polishing meshes when your loads are guesses

Most fatigue surprises I see trace back to bad loads, not bad models. We burn weeks on mesh refinement, then feed the solver a generic sine-on-random or legacy spectrum that ignores how the hardware actually vibrates.

Recent example: a gearbox mount bracket that our CAE said had plenty of life. Flight test found cracks around 120 hours. Post-test data showed narrowband peaks from gear mesh sidebands and a control-loop tone riding on the random. Our shaker spec smeared that content, and we modeled the joint as linear. Once we used measured spectra, added joint slip and preload scatter, the crack site and life lined up.

Lesson for me: measure early, even if it is taxi tests and ground runs. Identify mode shapes and damping, model joints as nonlinear, and bound friction. Then refine the mesh if it still matters. The solver is rarely the weak link.

How are you building realistic spectra and joint models into early predictions? Any lightweight methods you trust before you have full flight data?