Machine-learning ultrasonic detection of fatigue cracks: a physics-based simulation study

Fatigue cracks in microtextured titanium alloys are hard to find by ultrasound. Backscatter from alpha colonies, modulated by microtextured regions, mimics a small flaw, and crack faces held in contact by residual stress reflect far less energy than their area suggests. This paper is a simulation study. It builds a forward model of pulse-echo inspection whose constants are those of mill-annealed Ti-6Al-4V, verified against closed-form solutions on 15 quantities, and uses it to characterise what a learned detector can and cannot deliver under that model. The model solves a bilinear contact law at the crack interface, coupling the linear and nonlinear scattering paths. The detector reads the drive and second-harmonic bands through a small dilated residual encoder trained on simulated records only, and converts the score into a decision by split conformal calibration, whose false-call guarantee holds under exchangeability and is not claimed under distribution shift. On simulated test specimens with crack half-lengths from 0.05 to 1.0 mm, the encoder calls 73 of 84 low-contact cracks with no false call on 180 defect-free specimens (95% upper bound 0.017), and its fitted half-length at 90% detection is 0.138 mm (bootstrap interval 0.088–0.181 mm). A second-harmonic peak with no learning has the highest part-level area under the curve, the four leading methods are not separable on that measure, and an input ablation shows that an encoder reading the second-harmonic envelope alone matches or exceeds the full model, so the harmonic band carries the performance; over five training seeds the low-contact detection count spans 73 to 82 of 84. The encoder adds decision efficiency at the operating point, nine additional low-contact detections against one lost relative to the harmonic peak (paired test p = 0.02 ), not information absent from the measurement. Detection is confined to contact fractions between 0.05 and 0.30. The decision layer alone is then applied, as a methodological case study, to measured guided-wave data from an aluminium plate cycled to 100,000 cycles, where the second-harmonic channel gives 4 false calls in 370 angle tests and the drive band 9. No experimental validation on Ti-6Al-4V or on any other titanium alloy is presented, the only measured data are aluminium, and no qualification is claimed.