Simultaneous Improvement of Multiple Electrical Characteristics in Ultrathin a-IGZO TFTs Using Machine Learning Optimization

Abstract

While IGZO is emerging as a promising channel material to address the scaling limitations of conventional silicon-based DRAM, its application in next-generation 3D DRAM requires further advancements in achieving ultrathin structures and excellent performance tailored to DRAM characteristics. Specifically, optimizing process variables is essential for enhancing mobility in ultrathin structures, where mobility tends to degrade significantly, while maintaining a constant threshold voltage, a task that is both experimentally intensive and resource-demanding. This study employed multi-objective Bayesian optimization (MOBO) machine learning (ML) to simultaneously optimize multiple electrical objectives, aiming to achieve high mobility and a near-zero threshold voltage for ultrathin IGZO thin-film transistors (TFTs) under complex sputtering conditions, involving a wide range of possible combinations of Ar gas flow, sputtering power, and working pressure. Integrating empirical insights and expert knowledge into feature extraction, the MOBO approach leveraged human-driven expertise to optimize field-effect mobility and threshold voltage within the solution space. With ML assistance, a Pareto-optimal front was constructed to visualize trade-offs, achieving high field-effect mobility of 33.1 cm²/V·s and near-zero threshold voltage of −0.05 V at a 7.47 nm channel thickness. This approach is expected advance next-generation semiconductor technologies, offering exceptional gains in both efficiency and performance.