Predicting the DC Bias and Optimizing BER and PAPR in DCO-OFDM: An Explainable Machine Learning Approach

This paper proposes an explainable machine-learning framework for adaptive DC-bias selection and optimization in DC-biased orthogonal frequency division multiplexing (DCO-OFDM) systems. A large-scale synthetic dataset comprising 20,000 OFDM instances is generated using a physics-consistent DCO-OFDM signal model by sweeping the bias scaling factor. Each instance is labeled with the minimum DC bias that satisfies a target reliability constraint, while the corresponding minimum required target signal-to-noise ratio (TSNR) and peak-to- average power ratio (PAPR) are also recorded. Using only compact signal statistics together with system parameters, a LightGBM regressor accurately predicts the optimal DC bias under a leak-safe evaluation protocol that includes train-only preprocessing, fixed holdout testing, and multi-seed validation. The model achieves R² = 0.9946±0.0004 on in-distribution data and retains meaningful generalization performance under out-of-distribution settings. To enhance transparency, SHAP and LIME analyses are employed to interpret feature contributions. In a subsequent optimization stage, multi-output regression models are investigated to jointly predict DC bias V_DC(dB), PAPR, and TSNR. Among the evaluated models, the Gradient Boosting Regressor provides the best overall performance, achieving R² = 0.9614 for V_DC(dB), 0.9287 for PAPR, and 0.9570 for TSNR. A Pareto-based Optuna optimization then identifies nondominated operating points that capture the trade-offs among V_DC(dB), PAPR, and TSNR. Simulation results demonstrate consistent improvements in BER and PAPR over fixed-bias baselines, while preserving performance trends under deployment-relevant impairments, including LED front-end nonlinearity and static multipath optical channels.