On ultrasonic guided wave tomography using a decoder architecture for sparse sensor array data

We address defect mapping from sparse guided-wave measurements by learning to reconstruct 2-D velocity fields from eight travel-time paths. Our geometry-conditioned model is trained on a large synthetic dataset matched to a specific experimental setup. In simulation, reconstruction quality is highest when thinning overlaps the transmitter. On the experiment, the model recovers two corrosion regions; depth agrees (< 1mm), while x-y localisation is centre-biased. A fundamental limitation is the sim-to-real gap. Overall, eight paths suffice for coarse, depth-consistent maps with approximate estimation of x-y location.