Abstract
Physical modelling synthesis aims to generate audio from physical simulationsof vibrating structures. Thin elastic plates are a common model for drummembranes. Traditional numerical methods like finite differences and finiteelements offer high accuracy but are computationally demanding, limiting theiruse in real-time audio applications. This paper presents a comparative analysisof neural network-based approaches for solving the vibration of nonlinearelastic plates. We evaluate several state-of-the-art models, trained on shortsequences, for prediction of long sequences in an autoregressive fashion. Weshow some of the limitations of these models, and why is not enough to look atthe prediction error in the time domain. We discuss the implications forreal-time audio synthesis and propose future directions for improving neuralapproaches to model nonlinear vibration.