b'Figure 2. Schematic andtheoretical dispersion curvesspring compliances are introduced tousing dispersion curves obtained characterize the bond quality. First,through the forward computations. forward computations are conductedOnce trained, the neural network is for a wide range of the compliance ofready to solve the inverse problem, the spring layer. This will generate ai.e., using a set of experimentally training set of dispersion curves. Toobtained dispersion curves to obtain account for uncertainty and noisethe corresponding compliance of the in the experimental data, Gaussianspring layer. Finally, the performance random noise will be introduced inof the network is assessed based on the forward computation of disper- loss and accuracy.sion curves. Second, a convolution neural network (CNN) architecture will be developed and trained 158 2023|AFC ACCOMPLISHMENTS'