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

The paper introduces a novel framework for neural engine sound synthesis titled Physics-Informed Neural Engine Sound Modeling with Differentiable Pulse-Train Synthesis. This work is significant because it moves beyond purely data-driven spectral modeling by integrating explicit domain knowledge—such as physical pulse dynamics and mechanical sequencing—directly into the neural architecture, thereby achieving superior spectral reconstruction and harmonic accuracy while providing interpretable parameters that map directly to real-world engine mechanics.

Physics-Informed Inductive Biases

The core innovation of the PTR architecture lies in its use of physics-informed inductive biases, which guide the model's learning process away from purely spectral targets. The authors demonstrate that cycle-locked, decay-constrained pulse parameterization acts as a stronger inductive bias for impulsive periodic sources than free per-harmonic amplitude modeling. This approach focuses on modeling the underlying physical mechanisms—the pressure pulses generated during combustion—rather than merely reconstructing the resulting sound spectrum. PTR's structure is designed to capture complex mechanical transitions, such as combustion events becoming intermittent and resume synchronization upon re-engagement during clutch disengagement, a behavior arising from the pulse-train architecture itself.

Architectural Integration of Domain Knowledge

PTR achieves its enhanced performance by integrating domain knowledge at multiple architectural levels. The model's structure is explicitly designed to mimic physical processes:

• Parameterized Pressure Pulse Generation: This component models the initial impulsive energy source of combustion.

• Firing-Order Sequencing: This mechanism imposes a temporal constraint reflecting the mechanical firing order of the engine cylinders.

• Differentiable Karplus-Strong Resonators: These resonators are used to model complex acoustic phenomena, specifically producing convincing exhaustpipe resonances.

Quantitative Performance Gains Over Baselines

The quantitative validation results confirm PTR's superiority compared to a Harmonic-Plus-Noise (HPN) baseline. Across all tested datasets, PTR consistently outperforms the baseline, achieving improvements ranging from 3.8% to 7.6% in total validation loss. Specifically, the mean performance metrics show a 5.7% total loss reduction and 21% improvement in harmonic reconstruction. Furthermore, the model demonstrates robustness by maintaining consistent performance across diverse engine configurations, including successful generalization to Dataset A despite its architectural prior being based on an V8 firing-order setup.

Qualitative Synthesis and Interpretability

Beyond quantitative metrics, the synthesis quality reveals an authentic engine character. The model naturally captures several complex acoustic behaviors:

• RPM-dependent harmonicity: The sound's tonal content changes correctly with rotational speed.

• Load-dependent noise coupling: The relationship between mechanical load and background noise is accurately modeled.

• Natural Articulation Gradient: At low RPM, individual combustion events are clearly audible, while at high RPM, they smoothly blend into dense harmonic textures.

This integration of domain knowledge not only improves spectral reconstruction but also exposes interpretable parameters that map directly to physical phenomena such as valve timing, phase modulation, and exhaust resonance, providing deep insight into how mechanical properties shape the final timbre.

The core innovations presented are not simply about synthesizing engine sounds; they represent a powerful framework for integrating physical domain knowledge and causality into deep generative models. These principles can be generalized to improve any complex, time-series generation task involving structured physical or mechanical processes (e.g., human speech production, fluid dynamics simulation, complex robotic movements).

Here are the specific improvements I recommend:

• Improvement: Develop a generalized architecture that replaces purely spectral loss functions (L STFT) with modules that model the underlying physical processes or causal chains generating the signal. Instead of predicting coefficients from an STFT, the model should predict parameters for known physical generators (e.g., pressure pulses, resonator decay rates, flow dynamics).

• Mechanism: Implement a Process-Constrained Latent Space. The latent space z must be structured such that components map directly to measurable physical quantities (e.g., valve lift magnitude, combustion pressure peak timing).

• Capability: The improved system can generate highly realistic, physically consistent time-series data in domains where the underlying physics is known (e.g., synthesizing realistic mechanical failures, simulating complex biological processes like gait cycles, or generating structured speech segments based on articulatory phonetics rather than just acoustic features).

• Improvement: Generalize the concept of separating and independently modeling different signal components (e.g., pulse-train, noise coupling, resonance). This requires a hierarchical, differentiable signal processing module.

• Mechanism: Implement a Multi-Modal Differentiable Synthesis Pipeline. The network should explicitly pass through specialized differentiable modules for:

1. Impulsive/Periodic Components (Pulse Generation): Modeling the timing and magnitude of discrete events (e.g., impacts, heartbeats).

2. Continuous Background Noise (Turbulence/Flow): Using dedicated noise generators conditioned on state variables (e.g., velocity, load).

3. Resonant/Decay Components: Utilizing differentiable digital signal processing filters (like the Karplus-Strong resonators) whose parameters are predicted by the network, allowing for accurate modeling of system decay and acoustic feedback loops.

• Capability: The system can synthesize complex audio or sensory data where different energy sources operate simultaneously and interact non-linearly, achieving unprecedented control over timbre and spectral evolution (e.g., simulating the combined sound of a jet engine at different power settings, or generating speech that correctly models vocal cord vibration and mouth articulation simultaneously).

• Improvement: Generalize the harmonic loss function (L Harmonic) from engine orders to any domain involving periodic energy peaks (e.g., musical harmony, rotational mechanics, orbital dynamics).

• Mechanism: Develop a Generalized Energy Constraint Module. This module must calculate the energy contribution around predicted fundamental frequencies and their integer multiples (n times f 0), ensuring that the model's output maintains predictable harmonic relationships consistent with known physics (e.g., Fourier analysis constraints, or rotational symmetry).

• Capability: The improved AI can generate structured signals that adhere to strict mathematical or physical periodicity rules, making it invaluable for:

• Music Synthesis: Ensuring generated musical motifs maintain accurate consonance and overtone series relationships.

• Vibration Analysis: Generating synthetic sensor data (e.g., from bridges or machinery) that accurately reflects predictable resonant frequencies and their decay rates.

• Improvement: Explicitly model mechanical or operational state changes that cause abrupt, non-linear shifts in signal generation (e.g., clutch disengagement, gear shifts).

• Mechanism: Integrate a State Transition Module (STM) into the architecture. This module takes discrete external states (e.g., Clutch Open, Throttle Off) and modulates the parameters of all upstream generators (pulse timing, noise coupling factor, resonator damping) simultaneously. The transition itself must be modeled as a controlled decay or ramp-up, rather than an abrupt cutoff.

• Capability: The system can generate highly believable narratives in time-series data—for instance, simulating the entire operational life cycle of a complex machine (startup to steady state to load change to failure) with physically plausible transitions between states.

Engine sounds originate from sequential exhaust pressure pulses rather than sustained harmonic oscillations. While neural synthesis methods typically aim to approximate the resulting spectral characteristics, we propose directly modeling the underlying pulse shapes and temporal structure. We present the Pulse-Train-Resonator (PTR) model, a differentiable synthesis architecture that generates engine audio as parameterized pulse trains aligned to engine firing patterns and propagates them through recursive Karplus-Strong resonators simulating exhaust acoustics. The architecture integrates physics-informed inductive biases including harmonic decay, thermodynamic pitch modulation, valve-dynamics envelopes, exhaust system resonances and derived engine operating modes such as throttle operation and Deceleration Fuel Cutoff (DFCO). Validated on three diverse engine types totaling 7.5 hours of audio, PTR achieves a 21% improvement in harmonic reconstruction and a 5.7% reduction in total loss over a harmonic-plus-noise baseline model, while providing interpretable parameters corresponding to physical phenomena. Complete code, model weights, and audio examples are openly available.