Due to massive developments in underwater applications, the scarcity of bandwidth resources for underwater communication has fuelled the need for more advanced technology in the field of underwater communication. Therefore, a high-speed underwater optical wireless (UOW) link, driven by 16-Quadrature Amplitude Modulation (16-QAM) modulation and polarization division multiplexing, has been proposed in this work to alleviate bandwidth constraints in underwater communication. But the dynamic nature of the underwater link makes it very challenging for conventional methods to model the complex and non-linear relation between key physical link parameters and performance metrics of the underwater link. To overcome this challenge, a data-driven machine learning framework, based on state-of-the-art machine learning models, has been developed in this work to predict the signal-to-noise ratio (SNR) of the proposed underwater link. The proposed underwater link can achieve an underwater range of 17.93 m while ensuring a Forward Error Correction (FEC) compatible bit error rate (BER) of 3.8 × 10−3. Furthermore, a novel closed-form expression of BER for 16-QAM modulation-based UOW link has been derived by considering the combined impact of underwater turbulence, absorption, and scattering. In addition to this, a physics-guided and machine learning-based neural network for adaptive modulation (PML-NN-AM) has been proposed in this work for switching modulation order as per the channel conditions, and it has demonstrated good agreement with the analytical adaptive modulation benchmark.