Neural-operator surrogates for outdoor acoustics using parameterized wind and terrain profiles

Outdoor sound propagation is strongly shaped by boundary layer meteorology and terrain, yet high-fidelity simulation remains too expensive for rapid scenario studies. This paper shows the development of neural-operator surrogates that map parameterized wind profiles and/or terrain profiles to acoustic fields over a region of interest, using training data from a two-dimensional linearized Euler equation finite-difference time-domain solver with Sobol'-sequence sampling. Two operator-learning families are evaluated: deep operator networks and Fourier neural operators (FNOs) on wind-only, terrain-only, and joint wind+terrain datasets spanning multiple profile families. Beyond direct prediction of complex pressure (and derived sound pressure level), also studied is residual learning relative to fixed reference environments (no-wind and flat-ground baselines) for single-input settings. The paper provides a controlled comparison of architectures and targets and identifies when residual learning improves conditioning and generalization by emphasizing environment-induced modifications of the acoustic field. In this controlled two-dimensional setting, FNOs gave the most reliable full-field fixed-grid predictions, with out-of-distribution mean absolute sound-pressure-level errors below 2 dB and simulation speedups on the order of 10 000. The remaining errors are concentrated near shadow boundaries, near-ground regions, and interference structures, identifying these regions as key targets for reliability assessment in future neural-operator surrogate models.