Acoustic inpainting of loudspeaker radiation balloons from sparse polar data using deep learning

Accurate characterization of loudspeaker radiation patterns is essential for system design and acoustic simulation, yet obtaining complete three-dimensional radiation balloons typically requires extensive and time-consuming measurements. This paper presents a deep-learning framework for the acoustic inpainting of high-resolution radiation balloons from a sparse set of only three polar measurement planes. By mapping non-uniform spherical coordinates onto a structured two-dimensional grid, we leverage a geo-informed convolutional U-Net architecture designed to learn high-dimensional spatial–spectral correlations from empirical radiation data within a curated dataset of measured professional loudspeakers. On unseen cabinets, the model achieves a global mean absolute error (MAE) of 1.24±0.07 dB, yielding an improvement of up to 0.65 dB over standard interpolation baselines. The practical utility of the model is validated through integration into professional simulation software, where the reconstructed patterns enable accurate line array modeling. This approach significantly reduces measurement complexity while maintaining the high fidelity required for advanced predictive acoustic engineering.