Large Language Models of Propeller Noise

Abstract

This paper presents a language-model-inspired data-driven surrogate for predicting tonal and broadband noise from an isolated propeller as a function of blade geometry, operating condition, and receiver position. The model predicts acoustic spectra directly on a fixed frequency grid and represents the spectrum through tokenized broadband and tonal branches, deterministic harmonic rendering, and recombination in linear acoustic power. This formulation retains harmonic-level spectral information while imposing physically motivated structure on the predicted tonal--broadband decomposition. To improve low-order tonal fidelity, the tonal pathway incorporates geometry-, operating-, and directivity-aware inductive structure. Measured spectra provide the primary supervision, while auxiliary Ffowcs Williams--Hawkings (FW--H) simulation data are used only in restricted supporting roles for broadband pretraining and directivity regularization. When evaluated on a held-out, unseen propeller, the model achieves competitive accuracy for both raw spectra and one-third-octave levels while preserving physically interpretable harmonic ordering and broadband trends. The results demonstrate the potential of language-model-inspired spectral tokenization for low-data propeller aeroacoustic surrogate modeling, while indicating that broader generalization will require larger and more systematically sampled measured or high-fidelity synthetic datasets.