Hybrid Machine Learning Framework for Joint Prediction of Window Mean and Bit Error Rate in SC-LDPC Decoding

Abstract

Modern low-latency communication systems increasingly rely on spatially coupled low-density parity-check (SC-LDPC) codes combined with windowed decoding (WD) to achieve high reliability with reduced latency and memory requirements. However, evaluating the intrinsic trade-off between decoding complexity and error performance typically measured by the average window iteration count (W_MEAN) and bit error rate (BER) still depends on computationally intensive Monte Carlo simulations, which limits rapid system optimization and real-time design exploration. To address this limitation, this paper proposes a hybrid machine learning framework for the joint, non-iterative prediction of W_MEAN and BER using a single set of code and channel parameters. A high-fidelity dataset is generated through extensive SC-LDPC windowed decoding simulations across varying window sizes, coupling lengths, and signal-to-noise ratio (SNR) conditions. Based on this dataset, a multi-output Random Forest Regressor is trained to exploit the shared underlying decoding dynamics that govern both computational complexity and decoding reliability. The proposed model achieves accurate simultaneous prediction of W_MEAN and BER, demonstrating strong generalization performance while significantly reducing system evaluation time compared to conventional simulation-based approaches. Feature-importance analysis further reveals the dominant influence of channel quality and coupling structure on both decoding effort and error performance. These results indicate that the proposed framework provides an effective surrogate modeling tool for fast design-space exploration and informed performance–complexity trade-off analysis. The methodology enables practical optimization of high-throughput SC-LDPC decoders and supports the development of adaptive and resource-efficient communication systems.