Ferroelectric Tunnel Junction Memristor Crossbar Array with Annealing Optimization for In-Memory Computing

Abstract

Ferroelectric tunnel junctions (FTJs) are promising nonvolatile memory devices for in-memory computing due to their nonfilamentary switching behavior, low operating current, and multilevel conductance programmability. In this article, the influence of postmetallization annealing (PMA) temperature on HZO-based FTJs is investigated for stable ferroelectric switching characteristics. This annealing process promotes orthorhombic phase crystallization and reduces interfacial traps, enabling saturated polarization from the initial cycle and endurance exceeding 104 cycles. A 48 × 48 FTJ crossbar array is fabricated to evaluate array-level functionality. Half-bias operation is successfully demonstrated, confirming that unselected cells remained stable without interference from neighboring cells. Uniform 3-bit conductance programming is achieved across all 2,304 cells, exhibiting clearly separated states. Furthermore, vector–matrix multiplication tests verified accurate array-level operation, yielding an output error deviation as low as 0.97%. Finally, the final fully connected layer of a CIFAR-10 classifier is implemented directly on the crossbar array, achieving an 88.3% classification accuracy, closely matching the 88.5% software baseline. These findings underscore the potential of HZO-based FTJ crossbar arrays as energy-efficient platforms for in-memory computing.