Deep Learning-Based Prediction of Far-Field Sound Directivity in Parametric Array Loudspeakers

Parametric array loudspeakers (PALs) enable highly directional sound projection through ultrasonic modulation and are widely used in applications such as immersive audio, assistive communication, and spatial sound control. Despite their potential, accurate far-field directivity modelling remains challenging due to nonlinear acoustic propagation and the complexity of sparse array configurations. This dissertation proposes an integrated framework combining analytical, numerical, and deep learning methods to address these challenges. A dual convolution model is developed to improve grating lobe prediction and nonlinear beam characterisation. To optimise array performance under varying steering angles, a particle swarm optimisation (PSO) method is introduced for transducer layout design. Full-wave acoustic simulations using the finite-element method (FEM) are employed to model propagation and validate the optimised arrays, demonstrating up to a 5 dB reduction in peak sidelobe level across the tested steering range. To overcome the computational burden of full-wave simulations, deep learning models, based on generative adversarial network (GAN) architectures are trained to infer far-field patterns from sparse near-field inputs. By integrating physics-informed modelling, numerical optimisation, and data-driven prediction, the proposed approach improves both design efficiency and accuracy. In particular, the GAN-based predictor achieves MAE = 0.0105 and RMSE = 0.0171 on peak-normalised directivity. Experimental measurements were conducted to validate the proposed models and inform their development, ensuring consistency between predicted and observed acoustic directivity patterns. These outcomes address critical limitations in existing PAL modelling techniques and support the feasibility of real-time, adaptive control of acoustic fields for advanced sound applications.