Abstract:This paper presents an innovative approach to reducing Peak-to-Average Power Ratio (PAPR) in Coherent Optical Orthogonal Frequency Division Multiplexing (CO-OFDM) systems. The proposed deep learning autoencoder-based model eliminates the computational complexity of existing PAPR reduction techniques, such as Selective Mapping (SLM), by leveraging a novel decoder architecture at the receiver. In addition, No side information is needed in our approach, unlike SLM which requires knowledge of the PAPR distribution. Simulation results demonstrate significant improvements in both PAPR reduction and Bit Error Rate (BER) performance compared to traditional techniques. It achieves error-free transmission with over 10 dB PAPR reduction compared to unmitigated and 1 dB gain over SLM technique. Furthermore, our approach exhibits robustness against noise and nonlinearity effects, enabling reliable transmission over optical channels with varying levels of impairment. The proposed technique has far-reaching implications for next-generation optical communication systems, where efficient PAPR reduction is crucial for ensuring reliable data transfer.