Model-free autoencoder for adaptive optimization in turbulent underwater optical wireless communication systems

Abstract

Underwater optical wireless communication (UOWC) is a promising candidate for extending future 6G networks toward underwater connectivity, enabling high-rate links in ocean environments. However, its performance can be constrained by dynamic channel impairments such as turbulence and link misalignment, which make reliable and adaptive transmission particularly challenging for next-generation wireless systems. In response, we propose a model-free autoencoder (AE) framework that enables adaptive, robust, and high-capacity UOWC without relying on explicit channel modeling. Based on deep learning, the framework embeds the physical UOWC channel directly in the forward pass while using a highly simplified backward pass for gradient propagation, allowing end-to-end system optimization under diverse and time-varying conditions. We apply the approach to multi-carrier geometric constellation shaping (GCS). Experiments in a bubble-induced turbulent channel demonstrate rapid adaptation to turbulence variations and clear generalized mutual information (GMI) improvements over model-based AE. Furthermore, extending the framework to a 2×2 MIMO link, we jointly train a CNN-based combiner with the GCS modules. Simulation results show that our method achieves up to 0.57 bits/IQ symbol GMI gain over conventional combining schemes and up to 3.8 dB transmit-power reduction at a target information rate. Experiments in a bubble-induced turbulent MIMO setup further validate its effectiveness and scalability. Our framework offers strong potential for future intelligent and self-adapting UOWC systems operating in complex underwater environments.