Edge-efficient AI for acoustic channel estimation in resource-constrained subsurface IoT systems

Resource-constrained subsurface Internet of Things (IoT) systems require artificial intelligence (AI) solutions that are not only accurate, but also efficient and robust under limited sensing, computation, and communication capabilities. In underground and other opaque heterogeneous environments, acoustic communication provides a practical alternative to radio-frequency signaling, yet reliable channel estimation remains challenging due to severe scattering, multipath propagation, and medium inhomogeneity. Existing physics-based approaches often rely on restrictive assumptions, while conventional data-driven models may lack communication awareness and struggle to generalize across complex environments. This paper proposes an edge-efficient AI framework for acoustic channel estimation in resource-constrained subsurface IoT systems. The proposed method employs a communication-aware neural acoustic field representation to reconstruct subsurface acoustic channels from limited observations. By adopting a complex-valued formulation, the model jointly captures amplitude and phase information, enabling accurate recovery of full-wave channel responses. In addition, a structured decomposition of direct and scattered fields introduces a physically meaningful inductive bias that improves reconstruction efficiency, robustness, and learning effectiveness in heterogeneous media. Experimental results in heterogeneous subsurface environments demonstrate that the proposed framework achieves up to 20.5 dB reconstruction SNR and supports reliable QPSK communication with a bit error rate below $$10^{-3}$$ at an operating SNR of 9 dB. These results show that the proposed approach provides an accurate and efficient AI-based solution for acoustic channel estimation under resource-constrained IoT settings, and offers a practical foundation for future edge-intelligent subsurface sensing, communication, and autonomous system deployment in extreme environments.