SonarKAN: Sonar Kolmogorov-Arnold Network for Disentangling Passive Sonar Signatures

Abstract Data-driven neural models for passive sonar can be accurate yet difficult to interpret and diagnose because propagation and source/receiver effects become entangled. We propose Physics-aligned Sonar Kolmogorov--Arnold Network (SonarKAN), an architecture that aligns Kolmogorov--Arnold networks with the additive structure of the passive sonar equation. By constraining the network to B-spline univariate pathways, SonarKAN explicitly disentangles transmission loss from source spectral signatures. Furthermore, we introduce a physics-informed ridge projection that injects spreading-loss priors to accelerate convergence. Controlled simulations demonstrate accurate component recovery and substantially improved robustness to label noise compared to a multilayer perceptron (MLP) baseline.