Title: Physics-Informed Neural Engine Sound Modeling with Differentiable Pulse-Train Synthesis

Abstract: Unlike sustained harmonic oscillations, engine noise is fundamentally generated by sequential exhaust pressure pulses. While conventional neural synthesis approaches typically focus on approximating the resulting spectral features, this study introduces a method that directly models the underlying pulse shapes and their temporal structure. We introduce the Pulse-Train-Resonator (PTR), a differentiable synthesis framework that produces engine audio by generating parameterized pulse trains synchronized with engine firing sequences and passing them through recursive Karplus-Strong resonators to simulate exhaust acoustics. This architecture incorporates physics-informed inductive biases, such as harmonic decay, thermodynamic pitch modulation, valve-dynamics envelopes, and exhaust system resonances, alongside derived operational modes like throttle control and Deceleration Fuel Cutoff (DFCO). Evaluated across three distinct engine types comprising 7.5 hours of audio data, the PTR model demonstrates a 21% enhancement in harmonic reconstruction and a 5.7% decrease in total loss compared to a harmonic-plus-noise baseline. Furthermore, it offers interpretable parameters linked to physical phenomena. The complete codebase, model weights, and audio examples are publicly accessible.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC