Data-Driven Aircraft Noise Modeling for Multidisciplinary Design and Trajectory Assessment

Licentiate thesis, 2026

Aircraft noise assessment is an important part of aircraft design and airport-operation planning. Conventional semi-empirical prediction tools can describe individual sources, propagation effects, and certification metrics with useful engineering fidelity, but their computational cost becomes restrictive when thousands of evaluations are required in multidisciplinary design optimization or large-scale studies of aircraft trajectories and operational procedures. This thesis investigates machine learning models for aircraft noise that reduce this cost while retaining high accuracy. The study first focuses on the noise prediction model in the engine conceptual design phase. A database produced with the Chalmers Noise Code (CHOICE) is used to map different turbofan design-point and operating variables to the Effective Perceived Noise Level(EPNL) of engine components and the total engine. A deep neural network is compared with a local K-nearest-neighbor model and with a stacking architecture in which support-vector regression combines the two base learners. All evaluated models achieve root mean square errors(RMSE) below 0.3 dB for every source and total noise on the test set. The stacking model is particularly useful when compressor or turbine stage-count changes introduce local discontinuities that are difficult for a globally smooth neural approximation to reproduce. The second part of this study extends from the noise of the engine conceptual design to two-dimensional noise maps of a specific aircraft model under different flight conditions and trajectories. A Density-Guided Physics-Informed Attention U-Net (DGPIA U-Net) is trained on CHOICE-generated take-off cases for an A320 class aircraft. Sparse trajectory variables are converted to continuous influence fields through adaptive Gaussian diffusion, and channel and spatial attention are used to emphasize informative physical variables and locations. Once trained, the model produces noise fields on the millisecond scale, two to three orders of magnitude faster than the convention aircraft noise prediction method. Together, the studies establish a cross-scale strategy for data driven aircraft noise modeling: low dimensional surrogates accelerate design space exploration at engine component level, while physics-informed convolutional models reconstruct spatial exposure patterns for trajectory assessment.