Neural Radiated-Noise Fields for Unmanned Underwater Vehicle Noise Spectrum Prediction in Three-Dimensional Scenes

Researchers have developed a neural radiated-noise field (NRNF) model that predicts underwater vehicle acoustic signatures across three-dimensional spaces using machine learning rather than traditional physics-based simulation. The model achieves 3.5 dB average prediction error in the 50-5000 Hz band and demonstrates improved spatial generalization through a learnable scene feature grid.

This research represents a significant advancement in underwater acoustic modeling by replacing computationally expensive physics-based simulations with a neural network approach. Traditional methods require extensive structural information and environmental parameters that are often unavailable or difficult to characterize, creating bottlenecks in acoustic signature prediction for unmanned underwater vehicles. The NRNF framework overcomes these limitations by learning continuous spatial-frequency relationships directly from empirical lake trial data, enabling real-time spectrum queries at arbitrary positions without explicit environmental modeling.

The innovation extends beyond simple acoustic prediction into the broader domain of neural field representations, a technique gaining traction across computer vision, robotics, and scientific computing. By encoding three-dimensional positions and frequencies sinusoidally and incorporating learnable environmental feature grids, the model captures both direct acoustic propagation and complex environmental effects simultaneously. The experimental design reveals performance variations across different extrapolation scenarios: horizontal extrapolation performs best while depth extrapolation proves most challenging, indicating the model captures horizontal acoustic properties more robustly than vertical stratification effects.

For defense and maritime applications, this technology enables rapid acoustic signature assessment for vehicle design evaluation and performance characterization without costly experimental programs or detailed structural specifications. The approach scales to real-world deployment scenarios where environmental conditions vary continuously. The demonstrated cross-run generalization capability suggests the model captures fundamental acoustic propagation physics rather than memorizing specific trial conditions, though the 3.5 dB error margin indicates room for refinement. Future development focusing on depth prediction stability and multi-frequency coherence modeling could enhance practical applicability across deeper operational ranges and broader frequency bands relevant to naval and commercial underwater systems.

• →Neural radiated-noise fields replace physics-based simulation with machine learning for three-dimensional acoustic spectrum prediction in unmanned underwater vehicles
• →The model achieves 3.5 dB average prediction error across 50-5000 Hz frequency band without requiring detailed structural or environmental information
• →Learnable scene feature grids significantly improve spatial generalization and prediction stability compared to models without environmental representation
• →Depth extrapolation presents the greatest technical challenge, suggesting neural models capture horizontal acoustic propagation better than vertical stratification effects
• →The approach enables rapid acoustic signature assessment for vehicle design evaluation with potential applications across defense, marine robotics, and commercial underwater systems