| Title |
Ultrasonic characterisation of grain orientations, elastic tensor and geometry of thick-section welds using deep learning |
| Authors |
Machado, Lucas Queiroz ; Blumensath, Thomas ; Samaitis, Vykintas ; Lowe, Michael ; Kalkowski, Michal |
| DOI |
10.58286/33615 |
| Full Text |
|
| Is Part of |
e-Journal of nondestructive testing: 14th European conference on non-destructive testing (ECNDT 2026), 15-19 June 2026, Verona, Italy.. Mayen : NDT.net. 2026, vol. 31, spec. iss. 7, p. 1-2.. ISSN 1435-4934 |
| Keywords [eng] |
inversion ; ultrasonic tomography ; austenitic welds ; array imaging |
| Abstract [eng] |
Ultrasonic inspection of thick-section austenitic welds is hindered by the effects of microstructural heterogeneity: columnar grains with spatially varying crystallographic orientation distort, split and attenuate the beam, making conventional imaging based on the straight-ray assumption unreliable. This work proposes a model-based deep neural network (DNN) inversion workflow to infer weld microstructure description from time-offlight (ToF) maps. Work to date focused on grain orientations; in this contribution, we extend it to weld geometry, array position, and the elastic tensor. Our analysis shows the importance of respective parameters for ToF prediction and the impact of their uncertainty on the inversion. Training data are generated synthetically using a fast shortest-ray-path (SRP) ray-tracing forward model. Two inverse tasks are addressed with network-based metamodels: (i) ToF-to-weld-map regression (fully-connected architecture) to estimate the orientation map ; and (ii) ToF-to-parameter regression (combined convolutional and fully-connected network). We compare the architectures and assess their generalisation using k-fold cross-validation. Grain-scale finite-element simulations and experimental FMC data from industry-relevant welds validate inversion results and demonstrate that the predicted weld maps/parameters enable improved Total Focusing Method reconstructions. |
| Published |
Mayen : NDT.net |
| Type |
Conference paper |
| Language |
English |
| Publication date |
2026 |
| CC license |
|
