Improvement of Dual-Stacked Oxide Thin Film Transistor Using Bayesian Optimization

Oxide semiconductor Thin Film Transistors (OS TFTs) have shown outstanding characteristics such as a low leakage current and a threshold voltage near 0 V, making them a promising technology. However, their field-effect mobility is lower compared to Si-based TFTs, which is considered a weakness. To overcome this challenge, extensive research has been conducted, leading to the development of OS TFTs. Among various study, Dual-stacked OS TFTs are characterized by layered structures composed of various oxide semiconductors, forming an active layer. The use of multiple materials in the active layer allows the advantages of various materials, but leads to complex physical and electrical interactions, complicating the fabrication process and the operational mechanism of these transistors. Optimizing the performance of dual stacked TFTs is, therefore, a substantial challenge. Recently, the application of machine learning in material science, especially in complex design spaces, has gained significant attention as an efficient approach. Among various algorithms, Gaussian process (GP)-based Bayesian optimization (BO) is recognized as a promising optimization algorithm that reduces trial-and-error. In this study, we aimed to optimize the performance of dual-stacked (IZO/IGZO) TFTs using BO. We simultaneously considered three key sputtering variables: Argon/Oxygen gas flow ratio (%), DC power (W), and working pressure (mTorr), all of which influence the overall performance of the TFTs. Since the performance of TFT cannot be defined by a single parameter, we employed a Figure of Merit (FoM), combining three representative output parameters: field-effect mobility, threshold voltage (V_th), and subthreshold swing (S.S.), to train the machine learning algorithm. Despite the complex experimental design, we successfully optimized the performance of the dual stacked (IZO/IGZO) TFTs using BO. By leveraging the active learning characteristics of BO, the sputtering conditions were optimized with the guidance of ML. The optimized TFT exhibited high performance, showing a mobility of 46.71 cm²V^-1s^-1, V_th of -0.10 V, and S.S. of 0.19 Vdec^-1. This resulted in a significantly improved field-effect mobility, more than twice that of conventional IGZO TFTs (20.1 cm²V^-1s^-1), without any degradation in other characteristics. This study demonstrates the feasibility of utilizing BO to fabricate high-performance TFTs under complex experimental conditions, involving numerous input variables (sputtering process variables) and output variables (performance parameters of TFTs). By leveraging the capabilities of ML, researchers can explore complex design spaces more efficiently, leading to the development of advanced materials and devices with improved performance. This study serves as a valuable example of how the integration of machine learning and materials science can drive innovation and progress in various technological domains.