MoS₂ Channel-Enhanced High-Density Charge Trap Flash Memory and Machine Learning-Assisted Sensing Methodologies for Memory-Centric Computing Systems

Driven by the shift of artificial intelligence (AI) workloads to edge devices, there is a growing demand for nonvolatile memory solutions that offer high-density, low-power consumption, and reliability. However, well-established 3D NAND Flash using polycrystalline Si (Poly-Si) channel encounters bottlenecks in increasing bit density due to short-channel effects and cell-current limitations. This study investigates molybdenum disulfide (MoS₂) as an alternative channel material for 3D NAND Flash cells. MoS₂’s low bandgap facilitates hole-injection-based erase, achieving a broader memory window at moderate voltages. Furthermore, adopting a low-k (≈2.2) tunneling layer improves the gate-coupling ratio, reducing program/erase voltages and enhancing reliability, with endurance up to 10⁴ cycles and retention of 10⁵ s. Comprehensive analyses, including thickness-dependent MoS₂ electrical measurements, temperature-dependent conduction studies, and Technology Computer-Aided Design (TCAD) simulations, elucidate the relationship between channel thickness and reliability metrics such as endurance and retention. Furthermore, deep reinforcement learning–driven Berkeley Short-channel IGFET Model (BSIM) parameter calibration enables seamless integration of the MoS₂ model with a fabricated page-buffer chip, allowing circuit-level verification of sensing margins. This methodology can be applicable to new channel materials for next-generation memory devices. These results demonstrate that MoS₂-based nonvolatile memory effectively meets high-density, low-power, and reliable storage needs, presenting a promising solution for AI-centric edge computing.