Evaluating the generalization of complex-weight neural networks over simulated Lamb wave responses from hollow spheres.

The advancement of complex-valued machine learning brings about new potential for the neglected phase information in acoustics research. A comparison between models that either ignore phase or represent complex numbers as pairs of reals and fully complex numbers yields results that indicate that complex-valued networks are as good as or better than the best fully real option, while having roughly half of the computer memory cost. This is performed using simulated partial wave responses (Lamb waves) for hollow spheres suspended in an infinite homogenous ocean. These spheres of varying thickness are classified based on material with fully connected networks applied to a frequency passband. For the hardest to classify subset of the generated data, there was comparable classification confidence and accuracy observed between the best-performing real network and the complex network. Perfect classification accuracy for unseen partial wave response data was achieved in some trained models, which suggests a disparity in minima in the gradient space and promotes further study into noise augmentation and convolution with simulated multipath channels.