When ML meets fatigue life, physics still wins

I trialed a regression model to estimate fatigue damage on a bracket from rainflow histograms and mean stress, trained on shaker coupon data. Cross-validation looked great.

Flight test broke it. Temperature drift shifted strain by +/- 50 microstrain; the model turned that bias into big damage deltas. The spectra also had long dwells and a few high spikes not in training. Shot peen residual stress and clamp-up changed local mean stress and crack closure, which the model could not see.

What worked: a simple baseline (Miner's rule with Walker or Goodman, notch Kt, residual stress estimate) plus a small ML correction for sequence and dwell effects. We added sanity checks: reject predictions when histogram KL divergence exceeds a threshold and require conformal prediction intervals below a set width. That kept us honest.

For those using ML on structural life or vibration, how are you gating predictions and flagging out-of-family data? Any practical out-of-distribution tests or uncertainty methods you trust on real flight data?