A Deep-Learning Surrogate Model for Predicting the Broadband Radiated Sound Power of Submerged Cylindrical Shells

Abstract

A residual multilayer perceptron (ResNetMLP) surrogate framework is presented to predict the broadband radiated sound power spectra of submerged circular cylindrical shells. The surrogate maps the shell mean radius, wall thickness, and longitudinal excitation position to the unit-force sound power spectra generated by a high-fidelity frequency-domain solver. The trained model is deployed within a diagonal power superposition scheme using equivalent nodal forces derived from an unsteady computational fluid dynamics surface pressure field. Comparing surrogate predictions with direct diagonal vibroacoustic calculations confirms the high predictive accuracy within the diagonal approximation. Crucially, a key limitation of sound power-based surrogates is highlighted: because acoustic power is a quadratic scalar quantity, it cannot capture phase-coherent load interaction, providing a clear rationale for future pressure-based surrogate formulations.