Stacking-Based Ensemble Learning for STT-MRAM Channel Detection

This paper presents a novel stacking-based ensemble learning framework for reliable channel detection in spin-transfer torque magnetic random-access memory (STT-MRAM) systems, particularly under challenging conditions such as unknown read offsets and asymmetric write errors. The proposed method integrates three base models, including multi-layer perceptron (MLP), convolutional neural network (CNN), and long short-term memory (LSTM), enabling the capture of a diverse range of error characteristics, including nonlinear distortions, burst errors, and read-disturb effects. A meta-model is trained to optimally combine the outputs of these base models, resulting in a more robust improvement in the system performance. Simulation results show that the ensemble detector achieves a bit error rate up to two orders of magnitude lower and significantly improves the frame error rate compared to conventional threshold-based and single-model neural detectors. When combined with error correction coding (ECC), the proposed system consistently outperforms prior approaches, maintaining high-performance decoding reliability even in high-offset and high-noise environments.