Locates delaminations — hidden separations between the layers of a carbon fiber laminate — from guided-wave ultrasonic inspection data.

Model card · Code and write-up

Click one of the example panels at the bottom of the app. Each is a held-out test panel the model never saw during training, and the name tells you how big the real flaw is — so you can check whether the model found it.

The middle output panel shows the preprocessing step that does most of the work: subtracting the pattern common to every panel, which leaves mostly the defect behind.

Dice 0.905 ± 0.053 across 71 held-out panels, against 0.134 for the best non-learned baseline. Recall 0.929, IoU 0.830.

The detection threshold defaults to 0.10, chosen to favor sensitivity over precision — in structural inspection a missed defect matters far more than a false alarm. Move the slider to see the trade-off.

Trained entirely on simulated data. A research demonstration, not an inspection tool. It has never been evaluated against a physical measurement and must not inform any real inspection decision. Full limitations are in the model card.

Kudela & Ijjeh (2021), Institute of Fluid Flow Machinery, Polish Academy of Sciences.